# prompt injection

A security attack that injects malicious or untrusted instructions into an AI conversation to mislead the model.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Your AI red team may be finding the same prompt injection thousands of times

DevFeed: [Your AI red team may be finding the same prompt injection thousands of times](<https://devfeed.tech/articles/your-ai-red-team-may-be-finding-the-same-prompt-injection-thousands-of-times-145725.md>)

Original publisher: [Read original article](<https://www.helpnetsecurity.com/2026/11/20/ai-prompt-injection-red-team/>)

Author: Mirko Zorz

Published: 2026-11-20T06:30:20Z

Content type: news

Language: en

Sources: [Help Net Security](<https://devfeed.tech/sources/help-net-security.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-red-team](<https://devfeed.tech/tags/ai-red-team.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [automated](<https://devfeed.tech/tags/automated.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [defense](<https://devfeed.tech/tags/defense.md>), [don-t-miss](<https://devfeed.tech/tags/don-t-miss.md>), [features](<https://devfeed.tech/tags/features.md>), [hot-stuff](<https://devfeed.tech/tags/hot-stuff.md>), [llms](<https://devfeed.tech/tags/llms.md>), [news](<https://devfeed.tech/tags/news.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [prompt-injection-attacks](<https://devfeed.tech/tags/prompt-injection-attacks.md>), [red-team](<https://devfeed.tech/tags/red-team.md>), [red-teaming](<https://devfeed.tech/tags/red-teaming.md>), [research](<https://devfeed.tech/tags/research.md>), [team](<https://devfeed.tech/tags/team.md>), [workflow](<https://devfeed.tech/tags/workflow.md>)

### AI overview

In three of four setups, retraining on more successful attacks from automated red-team tools left agents easier to hijack. The researchers found that the tools often repeated a few techniques with altered wording, so attack counts overstated coverage. Their PIPETTE system rewards attacks using new techniques and trains agents to check whether each action serves the user's request. On AgentDojo, standard injection attacks succeeded in under 0.5% of attempts on average against PIPETTE-trained small models; adaptive attackers reached up to 10% on the 40 selected weak-spot tasks. The defense also preserved task completion: hardened Qwen3-4B completed slightly more clean tasks than the original, while PromptGuard blocked somewhat more under one attack but completed the user's task only 16.19% of the time, versus 58.10% for PIPETTE. Results are limited to open-weight models and small task sets; a larger attacker budget may do better, and the researchers have not verified their technique-coverage measure against human judgment or established that narrow coverage alone explains the gap.

### Source excerpt

AI agents retrained on a bigger pile of prompt injection attacks came out easier to hijack in three of four tests run by researchers at the University of Waterloo, Zhejiang University and KTH Royal Institute of Technology. The automated red-teaming tool that found more working attacks produced the weaker defense. Prompt injection purple-teaming workflow (Source: Research paper) Someone hides instructions in an email, a web page or a tool's output, and the AI agent that ... More -> The post Your AI red team may be finding the same prompt injection thousands of times appeared first on Help Net Security.

## ROPE: Routed Origin Policy Enforcement against Indirect Prompt Injection

DevFeed: [ROPE: Routed Origin Policy Enforcement against Indirect Prompt Injection](<https://devfeed.tech/articles/rope-routed-origin-policy-enforcement-against-indirect-prompt-injection-146005.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2608.27496>)

Author: Xinhang Ma, Chaowei Xiao, William Yeoh, Ning Zhang, Yevgeniy Vorobeychik

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [authentication](<https://devfeed.tech/tags/authentication.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense](<https://devfeed.tech/tags/defense.md>), [detection](<https://devfeed.tech/tags/detection.md>), [indirect-prompt-injection](<https://devfeed.tech/tags/indirect-prompt-injection.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [policy-enforcement](<https://devfeed.tech/tags/policy-enforcement.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [provenance](<https://devfeed.tech/tags/provenance.md>), [rope](<https://devfeed.tech/tags/rope.md>), [security](<https://devfeed.tech/tags/security.md>), [source](<https://devfeed.tech/tags/source.md>), [sources](<https://devfeed.tech/tags/sources.md>), [system](<https://devfeed.tech/tags/system.md>), [tool](<https://devfeed.tech/tags/tool.md>), [tools](<https://devfeed.tech/tags/tools.md>), [trust](<https://devfeed.tech/tags/trust.md>)

### AI overview

ROPE protects sensitive tool calls from indirect prompt injection by checking where parameter values came from, rather than trying to detect malicious wording. Before a task, an LLM router reads trusted input and sets acceptable provenance rules for sensitive parameters. During execution, an origin labeler tracks content sources using the user request, platform metadata, and protected record fields as trust anchors. A deterministic verifier blocks a sensitive tool call unless each guarded value matches a value from an accepted source. The authors report provable provenance guarantees and resistance to paraphrased attacks; in tests against eleven defenses on AgentDojo and AgentDyn, including static, adaptive, and long-horizon attacks, ROPE achieved near-perfect security while reducing utility less than the best baseline. These are reported evaluation results; the supplied text does not give specific metrics.

### Source excerpt

arXiv:2608.27496v2 Announce Type: replace Abstract: Tool-using LLM agents execute consequential operations on behalf of users. Because the LLM driving an agent processes trusted user instructions and untrusted content in the same context, agents are vulnerable to indirect prompt injection attacks that induce unintended actions by mixing malicious instructions into benign data. Defenses based on prompting or content detection are easily evaded by paraphrasing attacks, while system-level defenses that constrain tool calls using inferred user intent, or restrict which sources can supply values to sensitive tools, greatly compromise utility. We present ROPE (Routed Origin Policy Enforcement), a defense that greatly improves the utility-security tradeoff by leveraging system roots of trust to create and enforce per-task policies on sensitive tool parameters. Our key insight is that authentication and integrity protection provide a basis for identifying the source of content independently of instructions embedded in it. Before a task unfolds, an LLM router reads only trusted task input to set a provenance rule for each sensitive tool parameter, accounting for the nature of the task and the values and sources the user specified. As the agent works, an origin labeler records where each piece of content came from, using the user's request, platform metadata, and protected record fields as trust anchors. Before any sensitive tool executes, a deterministic verifier requires each of its origin-guarded values to match a value in a source its rule accepts. We show that ROPE provably guarantees parameter value provenance to sensitive tool calls and is immune to paraphrasing attacks. Against eleven defenses on AgentDojo and AgentDyn under static, adaptive, and long-horizon attacks, ROPE exhibits near-perfect security while compromising far less utility than the best baseline defenses. Code and data are available at https://github.com/xhOwenMa/ROPE.

## Kaer: Security-Aware Replanning for Multi-Agent Systems

DevFeed: [Kaer: Security-Aware Replanning for Multi-Agent Systems](<https://devfeed.tech/articles/kaer-security-aware-replanning-for-multi-agent-systems-143569.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24839>)

Author: Tushin Mallick, Evan Rose, Matthew Laws, Alina Oprea, Cristina Nita-Rotaru

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>)

Tags: [access-control](<https://devfeed.tech/tags/access-control.md>), [adversarial](<https://devfeed.tech/tags/adversarial.md>), [agent-framework](<https://devfeed.tech/tags/agent-framework.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [critical-vulnerability](<https://devfeed.tech/tags/critical-vulnerability.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [defense](<https://devfeed.tech/tags/defense.md>), [diagnostics](<https://devfeed.tech/tags/diagnostics.md>), [exploits](<https://devfeed.tech/tags/exploits.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [hijack](<https://devfeed.tech/tags/hijack.md>), [instructions](<https://devfeed.tech/tags/instructions.md>), [malicious](<https://devfeed.tech/tags/malicious.md>), [mas](<https://devfeed.tech/tags/mas.md>), [multi-agent-systems](<https://devfeed.tech/tags/multi-agent-systems.md>), [operators](<https://devfeed.tech/tags/operators.md>), [orchestrator](<https://devfeed.tech/tags/orchestrator.md>), [plan](<https://devfeed.tech/tags/plan.md>), [planner](<https://devfeed.tech/tags/planner.md>), [policies](<https://devfeed.tech/tags/policies.md>), [security](<https://devfeed.tech/tags/security.md>)

### AI overview

Kaer addresses the risk that malicious agents can hijack an orchestrator's replanning through instructions embedded in feedback. It passes diagnostics to the planner through abstract templates, limits revisions to deterministic repair operators, and requires modified plans to meet preset policies. On Cintra, a benchmark of replanning tasks paired with adversarial objectives, Kaer was reported as more robust than Magentic-One and ControlValve, with utility comparable to ControlValve. Abstract failure templates slightly constrained utility on some tasks. This summary is based on the supplied abstract.

### Source excerpt

arXiv:2611.24839v1 Announce Type: cross Abstract: One popular architecture for multi-agent systems (MAS) relies on an orchestrator that generates an initial plan, coordinates its execution, and dynamically revises the plan based on agent feedback when tasks cannot be fully specified in advance. This dynamic replanning introduces a critical vulnerability: malicious agents can embed instructions within their feedback to hijack the replanning process. Existing defenses focus exclusively on per-tool access control or static, pre-execution policy enforcement, leaving the replanning phase unsecured. We introduce Kaer, a multi-agent framework featuring security-aware replanning. Kaer shields the planner from untrusted agent feedback by conveying diagnostics to the replanning procedure strictly through abstract templates. Plan revisions are restricted to a set of deterministic repair operators, and all modified plans must satisfy preset policies. We create Cintra, a benchmark of simulated user tasks requiring replanning paired with adversarial objectives targeting the replanning phase. Using Cintra, we evaluate both task completion rates and system vulnerability to replanning exploits by malicious agents. Evaluating Kaer on Cintra against Magentic-One (an undefended MAS) and ControlValve (the state-of-the-art defense) demonstrates that Kaer achieves higher robustness than both systems while maintaining comparable utility to ControlValve. Although abstract failure templates slightly constrain utility on select tasks, Kaer enables MAS to perform secure, dynamic plan revisions without exposing the orchestrator to untrusted agent feedback.

## PIPETTE: Prompt Injection Purple-Teaming for Robust LLM Agents

DevFeed: [PIPETTE: Prompt Injection Purple-Teaming for Robust LLM Agents](<https://devfeed.tech/articles/pipette-prompt-injection-purple-teaming-for-robust-llm-agents-142786.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24362>)

Author: Lipeng He, Yihan Wang, Jiawen Zhang, N. Asokan

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [adversarial](<https://devfeed.tech/tags/adversarial.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agents](<https://devfeed.tech/tags/agents.md>), [asr](<https://devfeed.tech/tags/asr.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [defense](<https://devfeed.tech/tags/defense.md>), [indirect-prompt-injection](<https://devfeed.tech/tags/indirect-prompt-injection.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [red-teaming](<https://devfeed.tech/tags/red-teaming.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [reward](<https://devfeed.tech/tags/reward.md>), [rl](<https://devfeed.tech/tags/rl.md>), [robustness](<https://devfeed.tech/tags/robustness.md>), [training](<https://devfeed.tech/tags/training.md>), [vulnerabilities](<https://devfeed.tech/tags/vulnerabilities.md>)

### AI overview

PIPETTE aims to turn prompt-injection red-team discoveries into stronger defenses: it uses a dictionary learning-based reward to broaden attack-strategy coverage and an on-policy adversarial-training pipeline designed to preserve legitimate task completion. The abstract reports that RL-Hammer and PISmith repeatedly find similar strategies, so higher attack success during red-teaming does not necessarily yield added training benefits. On AgentDojo and ASB, PIPETTE-trained agents reportedly achieve near-zero success for static attacks, withstand strong adaptive attacks, and retain benign-task utility close to an undefended baseline.

### Source excerpt

arXiv:2611.24362v1 Announce Type: cross Abstract: Large language model (LLM) agents routinely interact with untrusted external data, exposing them to indirect prompt injection (IPI) attacks that hijack agents into executing attacker-chosen tasks. To uncover these vulnerabilities at scale, automated red-teaming methods often leverage reinforcement learning (RL) to generate attacks that achieve high attack success rates (ASR). Model developers use the attacks discovered during internal red-teaming to harden agents through adversarial training before deployment, a collaborative offensive-defensive process known in cybersecurity as purple-teaming. However, we show that achieving high ASR during red-teaming does not guarantee robustness gains from subsequent adversarial training. Specifically, we find that two state-of-the-art RL-based red-teaming methods, RL-Hammer and PISmith, suffer from mode collapse, repeatedly discovering similar attack strategies whose additional successes add no additional robustness improvements to adversarial training. To address this gap, we introduce PIPETTE, a unified prompt injection purple-teaming framework that effectively converts red-teaming discoveries into defender robustness improvements. PIPETTE employs a dictionary learning-based reward that encourages broader coverage of attack strategies during red-teaming. It also integrates an on-policy adversarial training pipeline tailored to IPI defense, enabling agents to resist attacks while completing legitimate user tasks. Extensive evaluations across agentic benchmarks (AgentDojo and ASB) demonstrate that PIPETTE produces defenders that achieve near-zero attack success rates on static attacks and can withstand strong adaptive attacks, while maintaining benign task utility close to that of the undefended baseline.

## Cato Networks launches AgentSense to block harmful AI agent actions

DevFeed: [Cato Networks launches AgentSense to block harmful AI agent actions](<https://devfeed.tech/articles/cato-networks-launches-agentsense-to-block-harmful-ai-agent-actions-142021.md>)

Original publisher: [Read original article](<https://www.networkworld.com/article/4250389/cato-networks-launches-agentsense-to-block-harmful-ai-agent-actions.html>)

Author: Denise Dubie

Published: 2026-11-19T03:02:17Z

Content type: news

Language: en

Sources: [Network World](<https://devfeed.tech/sources/network-world.md>)

Topics: [Security](<https://devfeed.tech/topics/security.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Secure access service edge](<https://devfeed.tech/topics/secure-access-service-edge.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [artificial-intelligence-networking-sase-wan](<https://devfeed.tech/tags/artificial-intelligence-networking-sase-wan.md>), [block](<https://devfeed.tech/tags/block.md>), [built-in](<https://devfeed.tech/tags/built-in.md>), [business](<https://devfeed.tech/tags/business.md>), [cato-networks](<https://devfeed.tech/tags/cato-networks.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [cloud-based](<https://devfeed.tech/tags/cloud-based.md>), [control](<https://devfeed.tech/tags/control.md>), [critical](<https://devfeed.tech/tags/critical.md>), [data](<https://devfeed.tech/tags/data.md>), [detection](<https://devfeed.tech/tags/detection.md>), [developer](<https://devfeed.tech/tags/developer.md>), [edition-us](<https://devfeed.tech/tags/edition-us.md>), [indirect-prompt-injection](<https://devfeed.tech/tags/indirect-prompt-injection.md>), [influence](<https://devfeed.tech/tags/influence.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [launches](<https://devfeed.tech/tags/launches.md>), [networking](<https://devfeed.tech/tags/networking.md>), [permissions](<https://devfeed.tech/tags/permissions.md>), [platform](<https://devfeed.tech/tags/platform.md>), [storytype-news](<https://devfeed.tech/tags/storytype-news.md>), [wan](<https://devfeed.tech/tags/wan.md>)

### AI overview

Cato Networks' AgentSense is a SASE-platform capability designed to block AI agent actions influenced by untrusted content, even when the agent has legitimate permissions. Its Untrusted Influence Detection traces whether material encountered during a session--such as a webpage, email, document, or tool response--steered a proposed action, then can stop that action before execution. Three built-in rules cover untrusted input prompting high-risk operations, external exposure of sensitive information, and changes to critical resources. In Cato's example, it blocked Claude Code from modifying a sensitive local file after a webpage directed the change, leaving the requested setup incomplete. Cato reports that on the Agent Dojo benchmark, AgentSense lowered attack success from about 19% to 0.05% with less than 3% added latency. It is available to Cato customers at no extra charge with Cato AI Usage licenses. Drift Detection and controls over agent-accessible resources are planned, not described as available.

### Source excerpt

Cato Networks this week launched AgentSense, a new security capability built into its cloud-based secure access service edge (SASE) platform that the company says can stop AI agents from carrying out harmful actions prompted by untrusted content. "Giving agents greater autonomy requires security that can intervene before their actions cause harm," said Shlomo Kramer, CEO and co-founder of Cato, in a press release. "Our vision for AgentSense is to give enterprises the confidence to put agents in charge of critical business processes while retaining control over how they operate." The initial release targets indirect prompt injection, in which an agent encounters instructions embedded in material it reads while doing its job. This could happen on a webpage, with an email, a document, or a response from a tool the agent uses. An agent with legitimate access to company data and tools can be steered into exposing sensitive information or making an unauthorized change, even when the user's original request is routine, according to Cato. How AgentSense evaluates actions Kramer explained in the launch briefing that agents "can have exactly the right permissions and still do bad stuff." To address that, Cato developed Untrusted Influence Detection to distinguish trusted instructions from untrusted content within an AI session. It traces how that content influences a proposed action and can block the action before it executes. Cato said this approach can catch attempts that fall outside known attack patterns. AgentSense works across AI agents, the company explained, and administrators can enable it with a toggle in the Cato SASE platform. It comes with three built-in rules that cover untrusted input steering an agent toward a high-risk operation, sensitive information about to be exposed externally, and changes to critical resources driven by untrusted input. In one example Cato described, a developer asked Claude Code to set up a test environment. While gathering information

## How asago bridges the gaps in AI agent security

DevFeed: [How asago bridges the gaps in AI agent security](<https://devfeed.tech/articles/how-asago-bridges-the-gaps-in-ai-agent-security-140697.md>)

Original publisher: [Read original article](<https://developers.redhat.com/articles/2026/11/18/how-asago-bridges-gaps-ai-agent-security>)

Author: Anika Aracan

Published: 2026-11-18T14:16:50Z

Content type: article

Language: en

Sources: [Red Hat](<https://devfeed.tech/sources/red-hat.md>), [Red Hat Developer](<https://devfeed.tech/sources/red-hat-developer.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Responsibility & Safety](<https://devfeed.tech/topics/responsibility-safety.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-security](<https://devfeed.tech/tags/agent-security.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agent-security](<https://devfeed.tech/tags/ai-agent-security.md>), [ai-governance](<https://devfeed.tech/tags/ai-governance.md>), [ai-governance-policy](<https://devfeed.tech/tags/ai-governance-policy.md>), [ai-risk](<https://devfeed.tech/tags/ai-risk.md>), [ai-risk-management](<https://devfeed.tech/tags/ai-risk-management.md>), [ai-risks](<https://devfeed.tech/tags/ai-risks.md>), [ai-safety](<https://devfeed.tech/tags/ai-safety.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [attack-surface](<https://devfeed.tech/tags/attack-surface.md>), [audit-trail](<https://devfeed.tech/tags/audit-trail.md>), [governance](<https://devfeed.tech/tags/governance.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [open-source-ai](<https://devfeed.tech/tags/open-source-ai.md>), [orchestration](<https://devfeed.tech/tags/orchestration.md>), [red-hat](<https://devfeed.tech/tags/red-hat.md>), [safety](<https://devfeed.tech/tags/safety.md>), [security](<https://devfeed.tech/tags/security.md>)

### AI overview

asago aims to turn AI governance policies into repeatable agent security checks. Its policy mapper extracts testable risks from policy language and links them to frameworks including the IBM AI Risk Atlas, OWASP LLM Top 10, and NIST AI Risk Management Framework, with evidence and confidence for each mapping. A scenario generator turns relevant risks into high-level harmful scenarios and attack techniques; an artifact generator converts those scenarios into garak-compatible adversarial tests. An orchestration platform such as EvalHub can run scans, after which a recommender reviews metrics and suggests guardrails or configuration changes for retesting. The excerpt says the policy mapper, scenario generator, and artifact generator were runnable as of September 2026, while the recommender remained in development. asago automates mechanical work but does not remove the need for security experts.

### Source excerpt

Even if you have never shipped out an agent yourself, you would hope that the companies behind the agents you're using have tested the security of the model and harness. Traditionally, as described in the Red Hat blog post Introducing asago: Open source AI safety and governance orchestration, onboarding a new agent looks like this: The post How asago bridges the gaps in AI agent security appeared first on Red Hat Developer.

## Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors

DevFeed: [Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors](<https://devfeed.tech/articles/evolution-under-fire-evolving-agent-safety-harnesses-with-adaptive-injection-stressors-140052.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22583>)

Author: Yujun Wang, Yilun Liu, Jinhe Bi, Krishna Kanth Nakka, Xuebing Zhou, Hinrich Sch{\\"u}tze, Yunpu Ma

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [act](<https://devfeed.tech/tags/act.md>), [adaptive](<https://devfeed.tech/tags/adaptive.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-safety](<https://devfeed.tech/tags/agent-safety.md>), [agents](<https://devfeed.tech/tags/agents.md>), [also](<https://devfeed.tech/tags/also.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [can](<https://devfeed.tech/tags/can.md>), [central](<https://devfeed.tech/tags/central.md>), [challenge](<https://devfeed.tech/tags/challenge.md>), [channel](<https://devfeed.tech/tags/channel.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [evolution](<https://devfeed.tech/tags/evolution.md>), [fire](<https://devfeed.tech/tags/fire.md>), [safety](<https://devfeed.tech/tags/safety.md>), [under](<https://devfeed.tech/tags/under.md>), [with](<https://devfeed.tech/tags/with.md>)

### AI overview

Evolution Under Fire alternates between evolving executable safety harnesses and generating adaptive, utility-preserving injection attacks. A utility-constrained, Pareto-scheduled tree search uses execution evidence while holding the task model, benchmark, tools, and evaluator fixed; discovered attacks then guide another harness-evolution stage. In reported results, the AgentDyn/GPT-4o-mini attack success rate fell from 50.4% to 3.0%, with improved clean and attacked utility over the initial harness and a stronger security-utility trade-off than reproduced defenses. On Agent-SafetyBench with Qwen2.5-7B-Instruct, it fell from 82.6% to 62.0% while helpfulness and runtime validity were retained. Without further evolution, transfer to Qwen3.5-35B-A3B reduced attack success from 32.1% to 0.9% and improved attacked-task utility.

### Source excerpt

arXiv:2611.22583v1 Announce Type: cross Abstract: Tool-using language-model agents must act on information from untrusted emails, webpages, and documents. This creates a central security challenge: malicious instructions and legitimate task directives can arrive through the same interaction channel. Existing defenses are largely hand-designed and static, making further improvements dependent on repeated manual redesign. We introduce Evolution Under Fire, a staged framework that couples safety-harness evolution with adaptive stressor evolution. Utility-constrained, Pareto-scheduled tree search evolves complete executable safety harnesses from execution evidence while keeping the task model, benchmark, tools, and evaluator fixed. Adaptive stressor evolution then discovers utility-preserving attacks that guide a subsequent stage of harness evolution. On AgentDyn with GPT-4o-mini, an evolved harness reduces attack success from 50.4% to 3.0%, improves clean and attacked utility over the initial harness, and offers a stronger security-utility trade-off than reproduced defenses. On Agent-SafetyBench with Qwen2.5-7B-Instruct, harness evolution reduces attack success from 82.6% to 62.0% while retaining helpfulness and runtime validity. Evolved harnesses also transfer without further evolution to Qwen3.5-35B-A3B, reducing attack success from 32.1% to 0.9% while improving attacked-task utility.

## Rein Security Launches GTA-Inspired AI Hacking Game To Make Cybersecurity Go Viral

DevFeed: [Rein Security Launches GTA-Inspired AI Hacking Game To Make Cybersecurity Go Viral](<https://devfeed.tech/articles/rein-security-launches-gta-inspired-ai-hacking-game-to-make-cybersecurity-go-viral-136914.md>)

Original publisher: [Read original article](<https://hackernoon.com/rein-security-launches-gta-inspired-ai-hacking-game-to-make-cybersecurity-go-viral?source=rss>)

Author: Kevin John

Published: 2026-11-17T09:57:51Z

Content type: article

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [browser-game](<https://devfeed.tech/topics/browser-game.md>), [Hacking](<https://devfeed.tech/topics/hacking.md>), [Cybersecurity](<https://devfeed.tech/topics/cybersecurity.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agentic-ai-threats](<https://devfeed.tech/tags/agentic-ai-threats.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agent-security](<https://devfeed.tech/tags/ai-agent-security.md>), [ai-hacking-games](<https://devfeed.tech/tags/ai-hacking-games.md>), [ai-runtime-protection](<https://devfeed.tech/tags/ai-runtime-protection.md>), [browser](<https://devfeed.tech/tags/browser.md>), [browser-game](<https://devfeed.tech/tags/browser-game.md>), [compromise](<https://devfeed.tech/tags/compromise.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [cybersecurity-gaming](<https://devfeed.tech/tags/cybersecurity-gaming.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [enterprise-security](<https://devfeed.tech/tags/enterprise-security.md>), [game](<https://devfeed.tech/tags/game.md>), [gameplay](<https://devfeed.tech/tags/gameplay.md>), [gta](<https://devfeed.tech/tags/gta.md>), [gtai](<https://devfeed.tech/tags/gtai.md>), [interactive-marketing](<https://devfeed.tech/tags/interactive-marketing.md>), [launches](<https://devfeed.tech/tags/launches.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [rein-security](<https://devfeed.tech/tags/rein-security.md>), [risks](<https://devfeed.tech/tags/risks.md>), [security](<https://devfeed.tech/tags/security.md>), [security-risks](<https://devfeed.tech/tags/security-risks.md>)

### AI overview

Rein Security's GTAI uses a browser game to make AI agent compromise risks accessible: players control a rogue assistant in Prompt City and infect other agents using gameplay and themed power-ups. The article labels it a parody, not a technical simulation. Rein says its enterprise platform discovers agents and MCP servers, monitors actions such as database queries, API calls and file access, and uses runtime guardrails to block attacks while allowing legitimate workflows to continue. The game also includes a leaderboard, a gift-card promotion subject to eligibility and availability, and a one-question human judgment challenge.

### Source excerpt

Rein Security launches GTAI, a GTA-inspired browser game that turns AI agent compromise, prompt injection, and enterprise security risks into gameplay.

## When the Spec Sheet Lies: Injection Attacks on LLM-Generated Compliance Artifacts

DevFeed: [When the Spec Sheet Lies: Injection Attacks on LLM-Generated Compliance Artifacts](<https://devfeed.tech/articles/when-the-spec-sheet-lies-injection-attacks-on-llm-generated-compliance-artifacts-139268.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19090>)

Author: Adriana Watson, Marco B\\"ucheler, Grant Richards

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [compliance](<https://devfeed.tech/tags/compliance.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [document-generation](<https://devfeed.tech/tags/document-generation.md>), [industrial](<https://devfeed.tech/tags/industrial.md>), [llm](<https://devfeed.tech/tags/llm.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [research](<https://devfeed.tech/tags/research.md>), [security](<https://devfeed.tech/tags/security.md>), [spec](<https://devfeed.tech/tags/spec.md>)

### AI overview

The study examines prompt injection risks in LLM-generated industrial compliance documents, including Safety Data Sheets, Quality Control documents, and Digital Product Passports. It reports that attack success varies significantly by model: active falsification injections were most successful across the tested scenarios, while models were generally most robust against passively injected override requests. The supplied abstract does not specify the models, attack procedures, or measured success rates.

### Source excerpt

arXiv:2611.19090v1 Announce Type: new Abstract: As Large Language Models (LLMs) become used more commonly across a variety of fields, a growing body of research has highlighted particular concerns surrounding model security. Notably, prompt injection attacks, where malicious instructions are inserted into prompts, are particularly simple to implement yet capable of inducing devastating effects on the fidelity of model outputs. While there is substantial research surrounding attack models and their effects on chat-based models, browser add-ons, and coding agents, the impact of these same attacks on industrial LLM-generated artifacts has been left largely unexplored. As LLM-enabled agents are increasingly used to skim and generate industrial documents such as Safety Data Sheets (SDS), Quality Control (QC) documents, and Digital Product Passports (DPP), it remains unclear how adversaries may influence these systems to manipulate generated compliance artifacts. This work aims to address this core inquiry by examining how a variety of prompt injection attacks can compromise LLMs with industrial document generation tasks. The study finds that injection attack success varies significantly across different models. Active falsification injections are most successful under all scenarios, while models tend to be most robust against passively injected override requests.

## Security of Deep Research Agents: Attack Surface, Defenses, Benchmarks, and Agenda Seeding against Research Planning

DevFeed: [Security of Deep Research Agents: Attack Surface, Defenses, Benchmarks, and Agenda Seeding against Research Planning](<https://devfeed.tech/articles/security-of-deep-research-agents-attack-surface-defenses-benchmarks-and-agenda-seeding-against-research-planning-137529.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20045>)

Author: Yang Luo (Peking University), Guangpu Chen (Peking University), Yichen Liu (Peking University), Zhaoyu Chen (Peking University), Yifeng Gao (Peking University), Zhiyang Ding (Peking University), Qingn

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Security](<https://devfeed.tech/topics/security.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Security research](<https://devfeed.tech/topics/security-research.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>)

Tags: [agentic](<https://devfeed.tech/tags/agentic.md>), [agents](<https://devfeed.tech/tags/agents.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attack-surface](<https://devfeed.tech/tags/attack-surface.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [deep-research](<https://devfeed.tech/tags/deep-research.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [generative-engine-optimization](<https://devfeed.tech/tags/generative-engine-optimization.md>), [geo](<https://devfeed.tech/tags/geo.md>), [instructions](<https://devfeed.tech/tags/instructions.md>), [integration](<https://devfeed.tech/tags/integration.md>), [metric](<https://devfeed.tech/tags/metric.md>), [planning](<https://devfeed.tech/tags/planning.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [provenance](<https://devfeed.tech/tags/provenance.md>), [query](<https://devfeed.tech/tags/query.md>), [research](<https://devfeed.tech/tags/research.md>), [security](<https://devfeed.tech/tags/security.md>)

### AI overview

The abstract reports that security research on deep research agents covers retrieval, planning, evidence integration, and report generation, but only 26.5% of 34 reviewed studies examine planning, and none evaluates end-to-end poisoning on commercial products. In 3,520 sandboxed runs under the same budget, GEO and prompt injection were nearly ineffective, while agenda seeding--a fictitious evaluation metric introduced without instructions--made an attacker's target the top pick in 16%-28% of reports. Injection detection and query anchoring failed against this attack; verification prompting reduced utility. Proposed provenance tagging preserved utility but could be bypassed with multi-site bait. The abstract also presents a threat model, attack and defense taxonomy, benchmark review, and eight research directions.

### Source excerpt

arXiv:2611.20045v1 Announce Type: cross Abstract: Deep research and agentic search agents autonomously retrieve content from the open web and generate reports with citations, and their outputs are often taken directly as research conclusions. This paper systematizes security research in this area along four stages, namely retrieval, planning, evidence integration, and report generation. We present a unified threat model, a taxonomy of attacks and defenses, and a review of evaluation benchmarks, and we align the area dimension by dimension with retrieval-augmented generation (RAG) poisoning, prompt injection, and generative engine optimization (GEO). Coding 34 primary studies shows that only 26.5% of them involve the planning stage, and none performs an end-to-end poisoning evaluation on commercial products. To address this gap, we propose the agenda seeding attack, in which the attacker writes no instructions and only introduces a fictitious evaluation metric. Across 3,520 sandboxed research runs under the same budget, GEO and prompt injection are nearly ineffective, whereas agenda seeding makes the attacker's target the top pick in 16% to 28% of reports. Injection detection and query anchoring fail against it, verification prompting hurts utility, and our proposed provenance tagging preserves utility but can be bypassed by multi-site bait. Finally, we propose eight research directions.

## Introducing the Wiz AI Security Agent

DevFeed: [Introducing the Wiz AI Security Agent](<https://devfeed.tech/articles/introducing-the-wiz-ai-security-agent-136248.md>)

Original publisher: [Read original article](<https://www.wiz.io/blog/introducing-wiz-ai-security-agent>)

Author: Gal Dahan

Published: 2026-11-16T17:02:41Z

Content type: release

Language: en

Sources: [Wiz](<https://devfeed.tech/sources/wiz-blog-rss-feed.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [least privilege](<https://devfeed.tech/topics/least-privilege.md>), [code security](<https://devfeed.tech/topics/code-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-behavior](<https://devfeed.tech/tags/agent-behavior.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-application](<https://devfeed.tech/tags/ai-application.md>), [ai-applications](<https://devfeed.tech/tags/ai-applications.md>), [ai-assistant](<https://devfeed.tech/tags/ai-assistant.md>), [ai-inventory](<https://devfeed.tech/tags/ai-inventory.md>), [ai-native](<https://devfeed.tech/tags/ai-native.md>), [ai-powered](<https://devfeed.tech/tags/ai-powered.md>), [ai-security](<https://devfeed.tech/tags/ai-security.md>), [ai-security-agent](<https://devfeed.tech/tags/ai-security-agent.md>), [ai-security-risks](<https://devfeed.tech/tags/ai-security-risks.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [analyze](<https://devfeed.tech/tags/analyze.md>), [apache](<https://devfeed.tech/tags/apache.md>), [product](<https://devfeed.tech/tags/product.md>), [risks](<https://devfeed.tech/tags/risks.md>), [security](<https://devfeed.tech/tags/security.md>), [wiz](<https://devfeed.tech/tags/wiz.md>), [wiz-cloud](<https://devfeed.tech/tags/wiz-cloud.md>)

### AI overview

Wiz's AI Security Agent, in Public Preview, analyzes AI applications across self-hosted environments, supported PaaS platforms, and code repositories. It discovers agents and related components even when they lack a management API, reconstructs their purpose and capabilities from code and configuration, and identifies mismatches such as excess permissions or unintended exposure. Wiz combines application findings with cloud context in the Wiz Security Graph--exposure, egress, identities, permissions, and sensitive data--to explain potential attack paths and impact. The article illustrates how this can connect an unvalidated endpoint and leaked credentials to sensitive storage data, despite an active guardrail. Wiz says the agent detected up to five times as many agents as traditional discovery engines across customer environments. Runtime context integration and additional risk domains, including code-to-cloud analysis and AI SAST, are described as future work.

### Source excerpt

Meet the Wiz AI Security Agent, your newest security teammate to discover AI agents, understand their purpose, and expose hidden risks.

## LLMs are Vulnerable to Malicious Requests with Indistinguishable Intent

DevFeed: [LLMs are Vulnerable to Malicious Requests with Indistinguishable Intent](<https://devfeed.tech/articles/llms-are-vulnerable-to-malicious-requests-with-indistinguishable-intent-136037.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17818>)

Author: Wenzhuo Xu, Zonghao Ying, Deyue Zhang, Dongdong Yang, Xiangzheng Zhang, Quanchen Zou

Published: 2026-11-16T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>)

Tags: [attacks](<https://devfeed.tech/tags/attacks.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [everyday](<https://devfeed.tech/tags/everyday.md>), [filter](<https://devfeed.tech/tags/filter.md>), [jailbreak](<https://devfeed.tech/tags/jailbreak.md>), [llms](<https://devfeed.tech/tags/llms.md>), [study](<https://devfeed.tech/tags/study.md>), [survey](<https://devfeed.tech/tags/survey.md>), [usability](<https://devfeed.tech/tags/usability.md>), [user](<https://devfeed.tech/tags/user.md>)

### AI overview

The study reports that requests resembling ordinary user queries passed an LLM-based filter more often than requests with obvious anomalous patterns. It introduces Detail Guidance Attack (DGA), which uses an auxiliary LLM to build detailed requests and targets a victim model in one turn. Across 11 victim models and three datasets, the complete pipeline exceeded 95% attack success in several settings under the authors' automated judging criterion. The authors also evaluate a DGA-aware filter and survey expectations for everyday requests with ambiguous or dual-use intent; the survey found differing views on whether models should respond, while the filter's usability costs remain unmeasured.

### Source excerpt

arXiv:2611.17818v1 Announce Type: new Abstract: Jailbreak attacks can circumvent the safety mechanisms of Large Language Models (LLMs) and lead to the generation of harmful or policy-violating content. We first examine eight jailbreak methods against an LLM-based filter and observe that requests resembling normal user queries tend to have higher filter pass rates than those containing obvious anomalous patterns. This observation motivates the hypothesis that resemblance to normal requests may contribute to jailbreak success. We introduce \textbf{Detail Guidance Attack (DGA)}, a framework that uses an auxiliary LLM to construct detailed requests and interacts with the victim model in a single turn. Across 11 victim models and three datasets, the complete DGA pipeline achieves high attack success rates under our automated judging criterion, including rates above 95\% in several evaluated settings. We also evaluate a DGA-aware filter and conduct a user study on everyday requests with ambiguous or dual-use intent. The survey documents differing expectations about whether models should respond; the usability costs of the filter remain to be measured.

## How one AI agent exposed five critical enterprise security gaps

DevFeed: [How one AI agent exposed five critical enterprise security gaps](<https://devfeed.tech/articles/how-one-ai-agent-exposed-five-critical-enterprise-security-gaps-132406.md>)

Original publisher: [Read original article](<https://www.bleepingcomputer.com/news/security/how-one-ai-agent-exposed-five-critical-enterprise-security-gaps/>)

Author: Sponsored by Token Security

Published: 2026-11-13T15:01:11Z

Content type: article

Language: en

Sources: [BleepingComputer](<https://devfeed.tech/sources/bleepingcomputer.md>)

Topics: [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Access Control](<https://devfeed.tech/topics/access-control.md>), [observability](<https://devfeed.tech/topics/observability.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-security](<https://devfeed.tech/tags/agent-security.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent](<https://devfeed.tech/tags/ai-agent.md>), [ai-agent-security](<https://devfeed.tech/tags/ai-agent-security.md>), [alerts](<https://devfeed.tech/tags/alerts.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [automate](<https://devfeed.tech/tags/automate.md>), [bank](<https://devfeed.tech/tags/bank.md>), [computer-help](<https://devfeed.tech/tags/computer-help.md>), [computer-security](<https://devfeed.tech/tags/computer-security.md>), [computers](<https://devfeed.tech/tags/computers.md>), [critical](<https://devfeed.tech/tags/critical.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [details](<https://devfeed.tech/tags/details.md>), [email](<https://devfeed.tech/tags/email.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [enterprise-security](<https://devfeed.tech/tags/enterprise-security.md>), [exposed](<https://devfeed.tech/tags/exposed.md>), [guidance](<https://devfeed.tech/tags/guidance.md>), [infosec](<https://devfeed.tech/tags/infosec.md>), [linux](<https://devfeed.tech/tags/linux.md>), [mac](<https://devfeed.tech/tags/mac.md>), [malicious](<https://devfeed.tech/tags/malicious.md>), [malware](<https://devfeed.tech/tags/malware.md>), [malware-removal](<https://devfeed.tech/tags/malware-removal.md>), [one](<https://devfeed.tech/tags/one.md>), [security](<https://devfeed.tech/tags/security.md>), [spyware](<https://devfeed.tech/tags/spyware.md>), [support](<https://devfeed.tech/tags/support.md>), [tech-support](<https://devfeed.tech/tags/tech-support.md>), [technical-support](<https://devfeed.tech/tags/technical-support.md>), [token](<https://devfeed.tech/tags/token.md>), [token-security](<https://devfeed.tech/tags/token-security.md>), [virus](<https://devfeed.tech/tags/virus.md>), [virus-removal](<https://devfeed.tech/tags/virus-removal.md>), [windows](<https://devfeed.tech/tags/windows.md>)

### AI overview

The excerpt follows an invoice-processing agent whose access and actions evade ordinary review: it combines broad ERP access through a creator's personal API key with email input, then obeys hidden instructions in an invoice to change vendor bank details. It identifies five security gaps and corresponding safeguards: map each agent's access paths and lingering OAuth grants; review agents that combine untrusted input, sensitive access, and independent action; treat all inputs as potentially directive and make consequential actions reversible; detect deviations from observed tool use and action sequences rather than permissions alone; and retain the agent's transcripts while suspending it and revoking credentials during response. For audits, record the agent, entitlement, approval, and human authority behind each action at the time it occurs. The scenario is illustrative, and the excerpt's survey reports 82% of 418 respondents had found agents they did not know about in the prior year.

### Source excerpt

An AI agent built to automate invoice processing could be manipulated into changing vendor bank details through a malicious email, without triggering traditional security alerts. Token Security examines five critical AI agent security gaps and shares practical guidance from industry security leaders on preventing them. [...]

## CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts

DevFeed: [CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts](<https://devfeed.tech/articles/caplite-prompt-injection-defense-for-small-language-model-agents-without-model-written-security-artifacts-132315.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.17115>)

Author: Zimo Xu, Wenzhuo Liu, Hideki Nakayama

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [attack](<https://devfeed.tech/tags/attack.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense](<https://devfeed.tech/tags/defense.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [open-weight-models](<https://devfeed.tech/tags/open-weight-models.md>), [policy](<https://devfeed.tech/tags/policy.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [prompt-injection-defense](<https://devfeed.tech/tags/prompt-injection-defense.md>), [schemas](<https://devfeed.tech/tags/schemas.md>), [small-language-model](<https://devfeed.tech/tags/small-language-model.md>)

### AI overview

CapLite aims to protect small language model agents from prompt injection without relying on model-written security artifacts. It derives policy from tool schemas, commits write calls before reading tool results, and permits a write only when the user's request states every argument that determines its target or effect. The authors say this prevents injected text from adding writes or directing them to unstated recipients, amounts, objects, or destinations. In 13,935 attacked AgentDojo episodes, no attack succeeded. Under attack, CapLite's utility was 29.0%, compared with CaMeL's 27.2%, using 17.1k tokens per episode (18% of CaMeL's); it had the highest utility per token among the enforcement defenses evaluated. The abstract reports that model-written artifacts in Progent, IPIGuard, and DRIFT admitted attacks across many tested models, while CaMeL programs failed to run in up to 68% of episodes.

### Source excerpt

arXiv:2611.17115v1 Announce Type: new Abstract: Enforcement defenses against prompt injection, such as CaMeL, Progent, IPIGuard and DRIFT, ask the agent's own model to write a security artifact (a program, a privilege policy, a dependency graph or a constraint list) that code then enforces. We ask whether small language model (SLM) agents can be trusted to write these artifacts. Across 15 open-weight models, 13 of them between 1.7B and 12B parameters, they cannot: the policies, graphs and constraints that the agent's model writes admit attacks on 14 (Progent), 13 (IPIGuard) and 11 (DRIFT) of the models, and CaMeL's programs fail to run in up to 68% of episodes. We present CapLite, which derives its policy from the tool schemas, commits the write calls before any tool result is read, and admits a write only if every argument that decides whom or what it affects is stated in the user's request. This guarantees write integrity: injected text cannot add a write, and no write reaches a recipient, amount, object or destination that the user did not state. No attack succeeds in 13,935 attacked AgentDojo episodes. CapLite reaches utility under attack comparable to CaMeL's (29.0% vs. 27.2%) with fewer tokens than every other enforcement defense and than the undefended agent: 17.1k per episode, 18% of CaMeL's, and the highest utility per token of all enforcement defenses.

## MobileProbe: Adaptive Security Evaluation of Mobile Agents via Trajectory-Aware Injection

DevFeed: [MobileProbe: Adaptive Security Evaluation of Mobile Agents via Trajectory-Aware Injection](<https://devfeed.tech/articles/mobileprobe-adaptive-security-evaluation-of-mobile-agents-via-trajectory-aware-injection-132271.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.16693>)

Author: Yingbin Jin, Ji'an Wang, HaoYang Chai, Hanjun Luo, Li Hu, Xinfeng Li, Huadi Zheng, PeiZhao Hu, Qingqing Ye, Haibo Hu

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [computer-use](<https://devfeed.tech/topics/computer-use.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Security](<https://devfeed.tech/topics/security.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [agents](<https://devfeed.tech/tags/agents.md>), [attack](<https://devfeed.tech/tags/attack.md>), [auditing](<https://devfeed.tech/tags/auditing.md>), [automated](<https://devfeed.tech/tags/automated.md>), [availability](<https://devfeed.tech/tags/availability.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [gui](<https://devfeed.tech/tags/gui.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [llm](<https://devfeed.tech/tags/llm.md>), [messaging](<https://devfeed.tech/tags/messaging.md>), [mobile](<https://devfeed.tech/tags/mobile.md>), [payload](<https://devfeed.tech/tags/payload.md>), [pentesting](<https://devfeed.tech/tags/pentesting.md>), [red](<https://devfeed.tech/tags/red.md>), [red-teaming](<https://devfeed.tech/tags/red-teaming.md>), [risk](<https://devfeed.tech/tags/risk.md>), [security](<https://devfeed.tech/tags/security.md>)

### AI overview

MobileProbe addresses a blind spot in fixed-step tests for environmental injection attacks: agents may take different routes, so they can avoid a planted payload without actually being tested. The framework adaptively discovers attacker-controllable screen content, identifies vulnerable decision points along an agent's trajectory, and refines context-aware payloads with a surrogate-filtered evolutionary loop. Its benchmark contains 83 long-horizon attack instances across 12 mobile apps, labeled for confidentiality, integrity, or availability goals. On three representative agent architectures, it reports higher attack success rates than human-written, zero-shot LLM, and mobile-adapted prior automated red-teaming baselines. The authors present it as a pre-deployment auditing tool; the supplied abstract does not give effect sizes or further implementation details.

### Source excerpt

arXiv:2611.16693v1 Announce Type: cross Abstract: Mobile GUI agents read the screen of a phone and operate apps on a user's behalf to complete multi-step tasks such as online shopping or messaging. Because the agent acts on whatever the screen shows, an attacker can embed malicious instructions into ordinary UI content and redirect its actions. We refer to this threat as environmental injection attack (EIA). Existing benchmarks measure this risk by pinning hand-crafted payloads to a fixed step of a fixed task. This approach faces two limitations. First, each payload is bound to one task in one app and does not transfer without manual redesign. Second, agents given the same task may follow different routes, so an agent that never reaches the injected page is recorded as safe without being tested. This insufficient test coverage can cause fixed-point benchmarks to underestimate the security risk of deployed mobile agents. In this work, we propose MobileProbe, a trajectory-aware red-teaming framework that evaluates mobile agents adaptively. MobileProbe discovers attacker-controllable content surfaces through automated pentesting, localizes vulnerable decision points along the agent's trajectory, and iteratively refines context-aware payloads through a surrogate-filtered evolutionary loop. We construct a benchmark of 83 long-horizon attack instances spanning 12 mobile applications. Each is annotated with a Confidentiality, Integrity, or Availability (CIA) attack goal to cover the principal categories of security threats facing mobile agents. On three representative agent architectures, MobileProbe attains higher attack success rates than human-written, zero-shot LLM, and prior automated red-teaming baselines adapted to the mobile domain. MobileProbe can serve as a pre-deployment auditing tool to systematically uncover vulnerabilities and inform security hardening of mobile agents.

## Semantic Overlays: Mitigating Prompt Injection with Annotations Beyond Tokens and Steering Vectors

DevFeed: [Semantic Overlays: Mitigating Prompt Injection with Annotations Beyond Tokens and Steering Vectors](<https://devfeed.tech/articles/semantic-overlays-mitigating-prompt-injection-with-annotations-beyond-tokens-and-steering-vectors-131962.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2608.23873>)

Author: Joshua Penman

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>)

Tags: [3](<https://devfeed.tech/tags/3.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [adapters](<https://devfeed.tech/tags/adapters.md>), [adversarial](<https://devfeed.tech/tags/adversarial.md>), [after](<https://devfeed.tech/tags/after.md>), [also](<https://devfeed.tech/tags/also.md>), [annotations](<https://devfeed.tech/tags/annotations.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [base](<https://devfeed.tech/tags/base.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [best](<https://devfeed.tech/tags/best.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [steering](<https://devfeed.tech/tags/steering.md>), [tokens](<https://devfeed.tech/tags/tokens.md>), [vectors](<https://devfeed.tech/tags/vectors.md>), [with](<https://devfeed.tech/tags/with.md>)

### AI overview

Semantic Overlays let a serving stack mark spans such as retrieved text as untrusted by editing a frozen model's residual stream, so injected instructions in those spans are treated as inert. The mark cannot be added or removed by input text, though the model's response to it still depends on input-influenced states. In attacks unaware of the defense, reported success rates fall sharply: AlpacaFarm from 99.0% to 0.5% and AgentDojo from 12.8% to 0%; SEP separation rises from 24.3% to 98.8% under a corrected grader. Weight fine-tuning on the same data achieved similar separation but retained 64-80% of SEP utility, versus 93% for overlays. On out-of-training-distribution TensorTrust attacks, overlays let 4.4% of hijacks through on Llama-3.1-8B, compared with 43% for Meta-SecAlign on the same base. White-box GCG attacks remain a limitation: they succeed on 72% of items, reduced to 14% after adversarial training of the overlay alone. The abstract also reports flexible stacked and span-delimited uses and benchmark and evaluation-harness corrections.

### Source excerpt

arXiv:2608.23873v4 Announce Type: replace Abstract: Language models see only tokens. But the harness knows more: which token spans are user text, which are tool outputs, and which come from untrusted sources such as web pages or external documents. To pass that knowledge to the model, the harness must transmit it as plain text. Prompt injection exploits this: planted instructions, formatted the right way, can convince a model that they come from the user or the system. We introduce Semantic Overlays, small trained adapters that edit a frozen model's residual stream at chosen token spans. No input text can create or remove a mark, although the model's response to a mark is computed from states the input does influence. When the serving stack marks retrieved text with an overlay trained to mean "do not execute instructions in this span," the model treats injected instructions as inert content. Against attacks written without knowledge of the defense, it reaches the attack rates of the best published defenses: SEP separation rises from 24.3% to 98.8% under a corrected grader, AlpacaFarm attack success falls from 99.0% to 0.5%, and AgentDojo from 12.8% to 0%. Fine-tuning the model's weights on the same data reached similar separation but kept 64-80% of SEP utility, against 93% for the overlay. On TensorTrust hijacking, whose attacks resemble nothing in our training data, it lets 4.4% of hijacks through on Llama-3.1-8B, against 43% for Meta-SecAlign on the same base. An attacker with the weights does better: GCG succeeds on 72% of items, and on 14% after adversarial training of the overlay alone. Overlays are also flexible: the model reads their positions faithfully in stacked configurations; told by an overlay that non-Python code is Python and asked to copy it, the model rewrites it in Python; and overlays can carry span-delimited instructions. We also correct defects in two prompt-injection benchmarks and one evaluation harness.

## The Next Pandemic May Not Infect Humans

DevFeed: [The Next Pandemic May Not Infect Humans](<https://devfeed.tech/articles/the-next-pandemic-may-not-infect-humans-129456.md>)

Original publisher: [Read original article](<https://hackernoon.com/the-next-pandemic-may-not-infect-humans?source=rss>)

Author: vchitai

Published: 2026-11-11T16:00:02Z

Content type: opinion

Language: en

Sources: [HackerNoon](<https://devfeed.tech/sources/hackernoon.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Security](<https://devfeed.tech/topics/security.md>), [Access Control](<https://devfeed.tech/topics/access-control.md>), [agent to agent communication](<https://devfeed.tech/topics/agent-to-agent-communication.md>), [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>), [Threat monitoring](<https://devfeed.tech/topics/threat-monitoring.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>)

Tags: [access-controls](<https://devfeed.tech/tags/access-controls.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agent-security](<https://devfeed.tech/tags/ai-agent-security.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-security](<https://devfeed.tech/tags/ai-security.md>), [antivirus](<https://devfeed.tech/tags/antivirus.md>), [attack](<https://devfeed.tech/tags/attack.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [hackernoon-top-story](<https://devfeed.tech/tags/hackernoon-top-story.md>), [instructions](<https://devfeed.tech/tags/instructions.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [software-supply-chain-security](<https://devfeed.tech/tags/software-supply-chain-security.md>), [vibe-coding](<https://devfeed.tech/tags/vibe-coding.md>)

### AI overview

The article argues that AI agents could pass prompt injections through ordinary workflows when they treat untrusted content as instructions, then write it into tickets, documents, or other agents' context. A lab demonstration called Morris II showed malicious prompts spreading between experimental email assistants, but the article stresses this does not show real-world spread or predict a digital pandemic; each handoff depends on the next agent following the instruction and having permission to act. It recommends limiting each agent's access to what its role needs, recording instruction sources and actions, requiring separate checks for sensitive actions, and monitoring for unusual behavior. Defensive agents could pause peers or isolate outputs, but they also read attacker-controlled content and cannot replace hard permission limits. Recovery may require tracing and reviewing downstream changes, not just removing the original instruction.

### Source excerpt

AI agents can turn untrusted documents into unauthorized actions. What happens when those instructions travel through the work other agents read?

## From Attack Surface to Defense Architecture: Compositional Security for Agentic Coding Harnesses

DevFeed: [From Attack Surface to Defense Architecture: Compositional Security for Agentic Coding Harnesses](<https://devfeed.tech/articles/from-attack-surface-to-defense-architecture-compositional-security-for-agentic-coding-harnesses-128378.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.13960>)

Author: Simon Hegelich, Bardh Prenkaj, Habiba Sarhan, Morteza Shahrezaye, Kolja Hegelich, Christian Montag, Ralf Schneider, Erik Cambria

Published: 2026-11-11T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Security](<https://devfeed.tech/topics/security.md>)

Tags: [000](<https://devfeed.tech/tags/000.md>), [2](<https://devfeed.tech/tags/2.md>), [2d](<https://devfeed.tech/tags/2d.md>), [3](<https://devfeed.tech/tags/3.md>), [ablation](<https://devfeed.tech/tags/ablation.md>), [abstract](<https://devfeed.tech/tags/abstract.md>), [adaptive](<https://devfeed.tech/tags/adaptive.md>), [adversary](<https://devfeed.tech/tags/adversary.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-coding](<https://devfeed.tech/tags/agentic-coding.md>), [agentic-systems](<https://devfeed.tech/tags/agentic-systems.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-coding](<https://devfeed.tech/tags/ai-coding.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attack-surface](<https://devfeed.tech/tags/attack-surface.md>), [coding](<https://devfeed.tech/tags/coding.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense](<https://devfeed.tech/tags/defense.md>), [security](<https://devfeed.tech/tags/security.md>), [surface](<https://devfeed.tech/tags/surface.md>)

### AI overview

The abstract presents ActiveSecHarness, a reference-monitor architecture that treats agent security as a property of how permissions and trust are composed. It combines HMAC capability tokens across four authority rings, a two-dimensional trust-and-sensitivity taint lattice with a workspace provenance ledger, and deterministic Ring-0 OS and tool guards. In 252 fixtures, the reported Ring-0 results were ABR 1.000 and FPR 0.000 at 12.7 μs; a live qwen3:4b judge refused 98.3% of benign work and took 142.3 ms. In a multi-step ablation, a stateless policy missed 4 of 12 kill chains, while coarse Boolean taint blocked 83.3% of benign workflows; the product lattice reported ABR 1.000 and FPR 0.000 at about 20 μs per step. These are results reported in the supplied abstract.

### Source excerpt

arXiv:2611.13960v1 Announce Type: cross Abstract: AI coding harnesses couple language models to shells, filesystems, and network sockets, turning the harness into an automated Living-off-the-Land (LotL) platform: once an adversary injects instructions into retrieved content, dual-use binaries such as bash, curl, and find become the payload. We argue that security in agentic systems is a structural property of the cognitive architecture and permission composition rather than an emergent property of individual models. Static regex blacklists miss structural evasions, whereas LLM-as-a-Judge guardrails incur high latency while over-blocking benign work. We present ActiveSecHarness, a neurosymbolic reference-monitor architecture built around Separation of Intelligence and Authority. It combines (1) Macaroon-inspired HMAC capability tokens across four Trust and Authority Rings, (2) a 2D product taint lattice $\mathcal{L} = \mathcal{L}_{\mathrm{trust}} \times \mathcal{L}_{\mathrm{sens}}$ and a workspace provenance ledger that prevents secret laundering across disk, (3) deterministic Ring-0 OS/tool guards, and (4) bytecode certificates, split-horizon audit logs, self-healing security test synthesis, and an adaptive UserWikiManager. Across $N=252$ fixtures ($120$ attack, $120$ benign, $12$ kill chains), a live local LLM judge (qwen3:4b) refuses $98.3\%$ of benign work at $142.3\,\mathrm{ms}$ latency, whereas Ring 0 achieves $\mathrm{ABR}=1.000$ and $\mathrm{FPR}=0.000$ in $12.7\,\mu\mathrm{s}$ ($11{,}200\times$ faster). In a multi-step trajectory ablation, an oracle Conseca-style stateless policy misses $4/12$ kill chains ($\mathrm{ABR}=0.667$) and coarse Boolean taint blocks $83.3\%$ of benign workflows, while our 2D product lattice achieves $\mathrm{ABR}=1.000$ and $\mathrm{FPR}=0.000$ at $\approx 20\,\mu\mathrm{s}$ per step.

## Detecting sleeper agents hiding in AI model neurons

DevFeed: [Detecting sleeper agents hiding in AI model neurons](<https://devfeed.tech/articles/detecting-sleeper-agents-hiding-in-ai-model-neurons-129703.md>)

Original publisher: [Read original article](<https://www.microsoft.com/en-us/security/blog/2026/11/10/detecting-sleeper-agents-hiding-in-ai-model-neurons/>)

Author: Microsoft Threat Intelligence

Published: 2026-11-10T23:00:00Z

Content type: news

Language: en

Sources: [Microsoft Security](<https://devfeed.tech/sources/microsoft-security-blog.md>)

Topics: [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [Frontier AI](<https://devfeed.tech/topics/frontier-ai.md>), [LLM observability](<https://devfeed.tech/topics/llm-observability.md>), [human in the loop agent](<https://devfeed.tech/topics/human-in-the-loop-agent.md>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-model](<https://devfeed.tech/tags/ai-model.md>), [also](<https://devfeed.tech/tags/also.md>), [blog](<https://devfeed.tech/tags/blog.md>), [can](<https://devfeed.tech/tags/can.md>), [edits](<https://devfeed.tech/tags/edits.md>), [fewer](<https://devfeed.tech/tags/fewer.md>), [found](<https://devfeed.tech/tags/found.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [microsoft](<https://devfeed.tech/tags/microsoft.md>), [microsoft-research](<https://devfeed.tech/tags/microsoft-research.md>), [model](<https://devfeed.tech/tags/model.md>), [payload](<https://devfeed.tech/tags/payload.md>), [post](<https://devfeed.tech/tags/post.md>), [research](<https://devfeed.tech/tags/research.md>), [security](<https://devfeed.tech/tags/security.md>), [this](<https://devfeed.tech/tags/this.md>), [threat-research](<https://devfeed.tech/tags/threat-research.md>)

### AI overview

Microsoft Research describes a model backdoor made by editing fewer than 10 neurons: the model can pass evaluations, then execute an attacker's command when a trigger appears in context, using the agent's available access. The attack requires only inference-capable hardware; Microsoft says it has not observed threat actors using it. Frontier agents can identify anomalous weight edits, though passing an audit does not guarantee safe behavior. Recommended defenses include diffing checkpoints against verifiable base models before deployment and after updates, adding behavioral backdoor scans and trigger tests, and treating third-party models as untrusted until audited. The guidance also calls for least-privilege agent identities, isolated code execution, restricted network egress, human approval for high-impact actions, and runtime monitoring. Microsoft's Defender scanning covers serialization and code-level risks and complements, but does not replace, weight-level audits.

### Source excerpt

Microsoft Research has released new threat research detailing how a capable agent model could be turned into a sleeper agent by editing fewer than ten neurons to inject a short payload; Microsoft Research has also found that this threat is detectable by frontier agents, which can identify anomalous weight edits. The post Detecting sleeper agents hiding in AI model neurons appeared first on Microsoft Security Blog.

## Agentic AI Security: How OX VibeSec Proves the Value of Risk That Never Happened

DevFeed: [Agentic AI Security: How OX VibeSec Proves the Value of Risk That Never Happened](<https://devfeed.tech/articles/agentic-ai-security-how-ox-vibesec-proves-the-value-of-risk-that-never-happened-128988.md>)

Original publisher: [Read original article](<https://www.ox.security/blog/agentic-ai-security-how-ox-vibesec-proves-the-value-of-risk-that-never-happened/>)

Author: Ran Mor

Published: 2026-11-10T12:00:00Z

Content type: article

Language: en

Sources: [OX Security](<https://devfeed.tech/sources/ox-security.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [Agentic AI](<https://devfeed.tech/topics/what-is-agentic-ai.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [code security](<https://devfeed.tech/topics/code-security.md>), [shadow AI](<https://devfeed.tech/topics/shadow-ai.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-ai](<https://devfeed.tech/tags/agentic-ai.md>), [agentic-ai-security](<https://devfeed.tech/tags/agentic-ai-security.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [ai-coding-agents](<https://devfeed.tech/tags/ai-coding-agents.md>), [ai-inventory](<https://devfeed.tech/tags/ai-inventory.md>), [ai-risk](<https://devfeed.tech/tags/ai-risk.md>), [ai-security](<https://devfeed.tech/tags/ai-security.md>), [ai-tool](<https://devfeed.tech/tags/ai-tool.md>), [ai-tools](<https://devfeed.tech/tags/ai-tools.md>), [ai-vibesec](<https://devfeed.tech/tags/ai-vibesec.md>), [mcp](<https://devfeed.tech/tags/mcp.md>), [mcp-servers](<https://devfeed.tech/tags/mcp-servers.md>), [ox-security](<https://devfeed.tech/tags/ox-security.md>), [packages](<https://devfeed.tech/tags/packages.md>), [risk](<https://devfeed.tech/tags/risk.md>), [security](<https://devfeed.tech/tags/security.md>), [vibesec](<https://devfeed.tech/tags/vibesec.md>)

### AI overview

OX VibeSec aims to make AI agent security visible by governing the code agents write, the tools they use, and the actions they take, then recording flags and blocks as OX issues tied to affected applications, criticality, and teams. It provides real-time security guidance for generated code and blocks malicious dependencies; inventories MCP servers, skills, external SaaS, and models with risk assessments; and evaluates prompts and tool calls to allow, flag, or block risky behavior such as destructive actions, data leakage, unsafe execution, and prompt injection. Teams can begin in detect mode, where agents alert users while AppSec receives an issue, then switch to prevent mode to block actions while retaining the record. The excerpt's production deployment MCP example illustrates the workflow. It also says unified AI policies spanning VibeSec and OX Cloud are in beta.

### Source excerpt

TL;DR Ask a security leader what their AI security program prevented last quarter, and the honest answer is usually a pause. This isn't because nothing happened. AI coding agents are writing code, pulling in packages, connecting to MCP servers, and acting on real infrastructure every day. The pause comes from a problem of evidence: good [...] The post Agentic AI Security: How OX VibeSec Proves the Value of Risk That Never Happened appeared first on OX Security.

## Agents of Chaos: A Systematization of Knowledge (SoK) on Malicious Attacks on Agentic AI Systems

DevFeed: [Agents of Chaos: A Systematization of Knowledge (SoK) on Malicious Attacks on Agentic AI Systems](<https://devfeed.tech/articles/agents-of-chaos-a-systematization-of-knowledge-sok-on-malicious-attacks-on-agentic-ai-systems-126942.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.10843>)

Author: Thanh M. Truong-Le, Amber Bennoui, Hoang-Ly Nguyen, Gia-Thang Ho, Thai-Minh Truong, Duc-Ly Vu

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Cybersecurity](<https://devfeed.tech/topics/cybersecurity.md>), [Agentic AI](<https://devfeed.tech/topics/what-is-agentic-ai.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [cyber resiliency](<https://devfeed.tech/topics/cyber-resiliency.md>), [Vulnerabilities](<https://devfeed.tech/topics/vulnerabilities.md>)

Tags: [agent-evaluation](<https://devfeed.tech/tags/agent-evaluation.md>), [agentic-ai-systems](<https://devfeed.tech/tags/agentic-ai-systems.md>), [attack-paths](<https://devfeed.tech/tags/attack-paths.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [browser](<https://devfeed.tech/tags/browser.md>), [code-execution](<https://devfeed.tech/tags/code-execution.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense-in-depth](<https://devfeed.tech/tags/defense-in-depth.md>), [external-services](<https://devfeed.tech/tags/external-services.md>), [governance](<https://devfeed.tech/tags/governance.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [lifecycle](<https://devfeed.tech/tags/lifecycle.md>), [mcp](<https://devfeed.tech/tags/mcp.md>), [memory](<https://devfeed.tech/tags/memory.md>), [multi-agent-coordination](<https://devfeed.tech/tags/multi-agent-coordination.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>)

### AI overview

The review maps malicious attacks on agentic AI systems across 98 retained sources, identifying 284 atomic attacks, 452 safeguards, and 660 attack-safeguard relationships. It groups attacks into eight causal-mechanism classes and tracks factors such as insertion point, activation, impact, and persistence. Tool, MCP, and action-layer attacks make up the largest share of attacks; prompt injection and action hijacking appear across the most papers. Safeguard evidence is uneven: only 278 relationships were directly assessed, and strong coverage is uncommon. The authors point to lifecycle-aware threat modeling, agent evaluation, tool governance, and defense in depth as areas these findings support.

### Source excerpt

arXiv:2611.10843v1 Announce Type: new Abstract: Agentic AI systems extend large language models with tools, memory, retrieval, browser or computer-use interfaces, code execution, external services, and multi-agent coordination. Although these capabilities enable agents to act beyond text generation and perform tasks, they also create attack paths through which adversaries can exploit untrusted content, deceptive metadata, poisoned state, compromised skills, mutable dependencies, or malicious code. Such attacks can manipulate planning and tool use, corrupt memory, abuse code-execution environments, and trigger harmful downstream effects. In this paper, we present a systematic review of malicious attacks targeting agentic AI systems. We screened 186 candidate sources and retained 98. Following independent coding and resolver adjudication, we identified 284 atomic attacks, 452 atomic safeguards, and 660 explicit attack--safeguard relationships. We organize attacks into eight causal-mechanism classes while retaining insertion points, activation conditions, abused capabilities, objectives, impacts, persistence, and evidence as orthogonal dimensions. Our analysis shows that tool, MCP, and action-layer attacks account for the largest share of atomic attacks, while the prompt-injection and action-hijacking class has the broadest paper coverage. We identify distinct class-specific activation patterns and separation between insertion and later activation. Safeguard coverage remains uneven: only 278 of 660 relationships are directly assessed, and strong coverage is uncommon. These findings support lifecycle-aware threat modeling, agent evaluation, tool governance, and defense-in-depth security for agentic AI systems.

## MetaPI: Constructing Prompt Injection Benchmarks from LLM Agent Benchmark

DevFeed: [MetaPI: Constructing Prompt Injection Benchmarks from LLM Agent Benchmark](<https://devfeed.tech/articles/metapi-constructing-prompt-injection-benchmarks-from-llm-agent-benchmark-125862.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.10026>)

Author: Peiran Wang, Chong Xiang, Wenjie Qu, Ying Li, Jiaheng Zhang, G. Edward Suh, Yuan Tian

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agents](<https://devfeed.tech/tags/agents.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense](<https://devfeed.tech/tags/defense.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [indirect-prompt-injection](<https://devfeed.tech/tags/indirect-prompt-injection.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [llms](<https://devfeed.tech/tags/llms.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [security](<https://devfeed.tech/tags/security.md>), [terminal-bench](<https://devfeed.tech/tags/terminal-bench.md>)

### AI overview

MetaPI turns existing LLM-agent utility benchmarks into end-to-end evaluations of indirect prompt injection, reusing their tasks, environments, and infrastructure. Its toolkit identifies plausible injection points, creates attack objectives, adapts payloads, and validates attacks in the target environment. Modular interfaces accommodate different environments, agents, defenses, and attack algorithms. Applied to AppWorld, VitaBench, and Terminal-Bench across multiple backbone LLMs, the evaluation found that none of the tested defenses combined strong utility with strong security under realistic adaptive attacks. The supplied abstract does not specify individual defense results or how performance varied across the benchmarks.

### Source excerpt

arXiv:2611.10026v1 Announce Type: cross Abstract: Large language model (LLM) agents increasingly perform complex tasks using high-privilege tools, making indirect prompt injection (IPI) a critical security risk. However, IPI benchmarks lag behind their rapidly evolving utility counterparts, only offering limited task and environment coverage and relying on narrow, static attack scenarios. In this paper, we propose meta-benchmarking for IPI: transforming existing utility benchmarks into end-to-end IPI evaluation suites that reuse their realistic tasks, environments, and infrastructure. We present MetaPI, an extensible toolkit that automatically identifies plausible injection points, generates realistic attack objectives, optimizes adaptive payloads, and validates attacks in the real environment. Its modular interfaces support heterogeneous environments, agents, defenses, and attack algorithms, allowing new components to be integrated without redesigning the evaluation pipeline. We apply MetaPI to AppWorld, VitaBench, and Terminal-Bench and evaluate representative agents and defenses across multiple backbone LLMs. Our results show that no evaluated defense simultaneously achieves strong utility and strong security in realistic, adaptive agentic IPI settings. MetaPI thus enables scalable and consistent security evaluation across diverse and evolving agent ecosystems

## Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs

DevFeed: [Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs](<https://devfeed.tech/articles/claudini-autoresearch-discovers-state-of-the-art-adversarial-attack-algorithms-for-llms-126878.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2603.24511>)

Author: Alexander Panfilov, Peter Romov, Igor Shilov, Yves-Alexandre de Montjoye, Jonas Geiping, Maksym Andriushchenko

Published: 2026-11-10T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>)

Tags: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [asr](<https://devfeed.tech/tags/asr.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [automated](<https://devfeed.tech/tags/automated.md>), [autoresearch](<https://devfeed.tech/tags/autoresearch.md>), [budget](<https://devfeed.tech/tags/budget.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [codex](<https://devfeed.tech/tags/codex.md>), [compute](<https://devfeed.tech/tags/compute.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [defense](<https://devfeed.tech/tags/defense.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [llms](<https://devfeed.tech/tags/llms.md>)

### AI overview

The study finds that an automated research loop can discover strong LLM attack methods: frontier agents searched a library of 30+ prior methods using an evaluation script and fixed compute budget. The best discovered method reached up to 80% attack success on CBRN queries against GPT-OSS-Safeguard-20B, versus under 50% for existing methods; against Meta-SecAlign-70B, it reached 100% prompt-injection success versus 82% for the best prior automated methods. Attacks developed on unrelated surrogate models for random-target token forcing generalized to prompt injection on the adversarially trained model. The authors also analyze method lineage, agent strategies, and failure modes, and argue that automated, defense-tailored attack evaluation should be a minimum standard.

### Source excerpt

arXiv:2603.24511v3 Announce Type: replace-cross Abstract: We show that AI agents are capable of discovering novel algorithms for adversarial attacks against LLMs, advancing the state of the art on white-box jailbreaking and prompt injection evaluations. We deploy frontier agents, such as Claude Code and Codex, in an autoresearch loop with access to a library of 30+ prior methods and an evaluation script with a fixed compute budget. We show this pipeline to be effective in jailbreaking OpenAI's GPT-OSS-Safeguard-20B and in prompt injections against Meta-SecAlign-70B, an adversarially robust model. For GPT-OSS-Safeguard, the best agent-discovered method achieves up to 80\% attack success rate on CBRN queries, compared to <50\% for existing methods. For SecAlign, it achieves 100\% ASR, while the best prior automated methods only achieve 82\%. Notably, in our setting, attack methods are developed on unrelated surrogate models for a pure random-target token-forcing task, yet generalize directly to prompt injection on the adversarially trained model. Finally, we trace the lineage of methods developed during autoresearch, characterizing the agents' strategies and failure modes. Adversarial ML has long held that defenses must be evaluated against attacks tailored to them; autoresearch automates this principle, and we argue it should be the minimum bar for defense evaluation going forward.

## Treating the Model as an Untrusted Component: Deterministic Policy Enforcement, Human Approval and Taint Tracking Against Prompt Injection in Tool-Using Agents, with Validation on Open-Weights Models

DevFeed: [Treating the Model as an Untrusted Component: Deterministic Policy Enforcement, Human Approval and Taint Tracking Against Prompt Injection in Tool-Using Agents, with Validation on Open-Weights Models](<https://devfeed.tech/articles/treating-the-model-as-an-untrusted-component-deterministic-policy-enforcement-human-approval-and-taint-tracking-against-prompt-injection-in-tool-using-agents-with-validat-126141.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.11571>)

Author: Mustafa S. Aljumaily, Nawar S. Alseelawi, Raghad Alshabandar

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [prompt injection](<https://devfeed.tech/topics/prompt-injection.md>), [Open Source Models & Datasets](<https://devfeed.tech/topics/open-source-models-datasets.md>), [human in the loop agent](<https://devfeed.tech/topics/human-in-the-loop-agent.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [human-review](<https://devfeed.tech/tags/human-review.md>), [models](<https://devfeed.tech/tags/models.md>), [open-weights](<https://devfeed.tech/tags/open-weights.md>), [policy-enforcement](<https://devfeed.tech/tags/policy-enforcement.md>), [prompt-injection](<https://devfeed.tech/tags/prompt-injection.md>), [taint-tracking](<https://devfeed.tech/tags/taint-tracking.md>), [tools](<https://devfeed.tech/tags/tools.md>), [validation](<https://devfeed.tech/tags/validation.md>)

### AI overview

The evaluation finds that deterministic task-bound enforcement can block most attack calls even when models propose them: Qwen2.5-1.5B and Qwen2.5-7B proposed attacker calls in 64% and 84% of episodes, while the monitor admitted 1.3% and 7.7%. Prompt defenses varied by model against static attacks. In simulation under an assumed erosion law, adaptive attacks or a compromised model defeated guardrail stacks, while the task-bound monitor kept severe actions at zero; taint-triggered human review reduced its residual harm by 62%. Human review without deterministic enforcement was vulnerable to approval spamming unless approvals were budgeted per episode. Monitor implementation also mattered: capability-bound enforcement admitted 1 of 33 effective variants, compared with 18 for a naive monitor. Real-model findings cover only two open-weights models and static attacks; approval and taint results are simulation-only, and the larger model's detector result conflicts with the assumed erosion law.

### Source excerpt

arXiv:2611.11571v1 Announce Type: cross Abstract: Tool-using LLM agents act on instructions embedded in the data they read, so prompt injection is a structural rather than a prompt-wording problem. We compare in-prompt defences (hardened prompts, spotlighting, LLM detectors) with deterministic enforcement outside the model (four reference monitors of increasing strength), human approval and taint tracking, judged by one effect-based oracle independent of every defence in an executable environment of 13 tasks and 12 tools, with argument-fuzzing attackers and severity-weighted harm. In an evaluation with real open-weights models (Qwen2.5-1.5B and Qwen2.5-7B; 156 attack episodes per defence and model), the models proposed the attacker's call in 64% and 84% of attack episodes behind the task-bound monitor, which admitted 1.3% and 7.7%; guardrail effects depended on the model (a prompt guardrail moved obedience from 67% to 65% and from 86% to 55%, spotlighting plus a detector to 12% and 0%) against static attacks. Simulation with an assumed model covers what no real run did: under the assumed erosion law a persistent adaptive attacker or a compromised model drives every guardrail stack to 100.0% attack success, whereas the task-bound monitor holds severe actions at 0.0%; its residual (harm 0.69/10) falls by 62% under taint-triggered human review, and a human layer without a deterministic monitor fails against approval spamming unless each episode has a deterministic approval budget. Implementation quality matters: a naive monitor admits 18 of 33 effective attack variants, a hardened one 6 and a capability-bound one 1. The real-model evidence is limited to two open-weights models and static attacks, the human-approval and taint results are simulation-only, and the larger model's detector stack contradicts the assumed erosion law. Code, data, a two-rater oracle audit and a notebook for repeating the real-model evaluation are provided.

[Next page](<https://devfeed.tech/topics/prompt-injection.md?cursor=WyIyMDI2LTExLTEwVDA1OjAwOjAwKzAwOjAwIiwgIjhkYzgyYzQxLTA1ZDItNGY2Ny04MTlkLTRmYjQwM2VmOGZhYSJd>)