# LLM security

LLM security is the practice of protecting large language models, their applications, and related systems from adversarial manipulation, misuse, and operational failure across development, deployment, inference, and application architecture.

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

## Forrester Recognizes Checkmarx in The Forrester Wave™: Agentic Development Security Tools, Q4 2026

DevFeed: [Forrester Recognizes Checkmarx in The Forrester Wave™: Agentic Development Security Tools, Q4 2026](<https://devfeed.tech/articles/forrester-recognizes-checkmarx-in-the-forrester-wavetm-agentic-development-security-tools-q4-2026-146870.md>)

Original publisher: [Read original article](<https://checkmarx.com/blog/checkmarx-forrester-wave-agentic-development-security-tools-q4-2026/>)

Author: Checkmarx Team

Published: 2026-11-20T17:03:20Z

Content type: article

Language: en

Sources: [Checkmarx](<https://devfeed.tech/sources/checkmarx.md>)

Topics: [code security](<https://devfeed.tech/topics/code-security.md>), [AI software supply chain security](<https://devfeed.tech/topics/ai-software-supply-chain-security.md>), [Static code analysis](<https://devfeed.tech/topics/static-code-analysis.md>), [AI-assisted coding](<https://devfeed.tech/topics/ai-assisted-coding.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [DevSecOps](<https://devfeed.tech/topics/devsecops.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-development](<https://devfeed.tech/tags/agentic-development.md>), [agentic-development-lifecycle](<https://devfeed.tech/tags/agentic-development-lifecycle.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-application](<https://devfeed.tech/tags/ai-application.md>), [ai-llm-tools-in-application-security](<https://devfeed.tech/tags/ai-llm-tools-in-application-security.md>), [analyst-reports](<https://devfeed.tech/tags/analyst-reports.md>), [application](<https://devfeed.tech/tags/application.md>), [aspm](<https://devfeed.tech/tags/aspm.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [blog](<https://devfeed.tech/tags/blog.md>), [checkmarx](<https://devfeed.tech/tags/checkmarx.md>), [code](<https://devfeed.tech/tags/code.md>), [development](<https://devfeed.tech/tags/development.md>), [forrester](<https://devfeed.tech/tags/forrester.md>), [forrester-wave](<https://devfeed.tech/tags/forrester-wave.md>), [q4](<https://devfeed.tech/tags/q4.md>), [remediation](<https://devfeed.tech/tags/remediation.md>), [security](<https://devfeed.tech/tags/security.md>), [security-tools](<https://devfeed.tech/tags/security-tools.md>), [tools](<https://devfeed.tech/tags/tools.md>)

### AI overview

Checkmarx says Forrester's Q4 2026 Agentic Development Security Tools evaluation gave it the highest possible scores for AI application flaws, autonomous proprietary code remediation, and ASPM for AI. The article frames agentic development security as a response to coding agents that generate code, choose dependencies, invoke tools, and execute tasks at machine speed. Checkmarx describes a three-part approach: combine rules-based and AI analysis, then consolidate findings and flag false positives; prevent issues through Developer Assist and prioritize existing findings by reachability, exploitability, business context, and compensating controls for guided-to-autonomous fixes; and correlate risk, assign owners, enforce policy gates, and deterministically govern models, MCP servers, and agents. It also says malicious and hallucinated packages are blocked in real time against a database of more than 550,000 known malicious packages. These are Checkmarx's descriptions of its capabilities; the text notes that Forrester does not endorse vendors or advise selection based on ratings.

### Source excerpt

Checkmarx received the highest score possible in the criteria of AI application flaws, autonomous proprietary code remediation, and ASPM for AI.

## Mapping the AI threat landscape: MITRE ATLAS comes to Elastic Security

DevFeed: [Mapping the AI threat landscape: MITRE ATLAS comes to Elastic Security](<https://devfeed.tech/articles/mapping-the-ai-threat-landscape-mitre-atlas-comes-to-elastic-security-146683.md>)

Original publisher: [Read original article](<https://www.elastic.co/security-labs/blog/mitre-atlas-elastic-security-detection-coverage>)

Author: Davis Plumlee

Published: 2026-11-20T16:30:00Z

Content type: article

Language: en

Sources: [Elastic Security Labs](<https://devfeed.tech/sources/elastic-security-labs.md>)

Topics: [ai security](<https://devfeed.tech/topics/ai-security.md>), [threat detection](<https://devfeed.tech/topics/threat-detection.md>), [Security](<https://devfeed.tech/topics/security.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [shadow AI](<https://devfeed.tech/topics/shadow-ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-automation](<https://devfeed.tech/tags/ai-automation.md>), [atlas](<https://devfeed.tech/tags/atlas.md>), [elastic](<https://devfeed.tech/tags/elastic.md>), [elastic-security](<https://devfeed.tech/tags/elastic-security.md>), [mitre](<https://devfeed.tech/tags/mitre.md>), [mitre-atlas](<https://devfeed.tech/tags/mitre-atlas.md>), [security](<https://devfeed.tech/tags/security.md>), [security-operations](<https://devfeed.tech/tags/security-operations.md>), [soc](<https://devfeed.tech/tags/soc.md>)

### AI overview

Elastic Security 9.6 adds MITRE ATLAS mappings alongside ATT&CK for prebuilt AI platform and GenAI detection rules, giving users a separate view of coverage for AI threats such as prompt injection, jailbreaks, stolen model API keys, and poisoned agent tools. Users can inspect coverage in a new table, add or edit both frameworks' mappings in rule creation workflows, and use Agent Builder to query covered threats or find rules for ATLAS techniques. A managed MITRE data service stores framework entities as Kibana Saved Objects and serves them through a data client and API, enabling queries and filtering and supporting multiple frameworks. Elastic says this may allow framework updates to follow the prebuilt-rule release cycle; the feature timing remains at its discretion.

### Source excerpt

Elastic Security 9.6 adds MITRE ATLAS to our detection rules and the tools you use to manage them, so you can see which AI attacks your rules already cover.

## Palo Alto Networks Expands into Enterprise AI with $4.5B Deal

DevFeed: [Palo Alto Networks Expands into Enterprise AI with $4.5B Deal](<https://devfeed.tech/articles/palo-alto-networks-expands-into-enterprise-ai-with-4-5b-deal-146662.md>)

Original publisher: [Read original article](<https://www.eetimes.com/palo-alto-networks-expands-into-enterprise-ai-with-4-5b-deal/>)

Author: Majeed Ahmad

Published: 2026-11-20T15:37:26Z

Content type: news

Language: en

Sources: [EE Times](<https://devfeed.tech/sources/ee-times.md>)

Topics: [AI Infrastructure](<https://devfeed.tech/topics/ai-infrastructure.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [acquisition](<https://devfeed.tech/tags/acquisition.md>), [advanced-ai-capabilities](<https://devfeed.tech/tags/advanced-ai-capabilities.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-accelerators](<https://devfeed.tech/tags/ai-accelerators.md>), [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [ai-infrastructure](<https://devfeed.tech/tags/ai-infrastructure.md>), [ai-models](<https://devfeed.tech/tags/ai-models.md>), [ai-systems](<https://devfeed.tech/tags/ai-systems.md>), [ai-tools](<https://devfeed.tech/tags/ai-tools.md>), [ai-training](<https://devfeed.tech/tags/ai-training.md>), [analog](<https://devfeed.tech/tags/analog.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [business](<https://devfeed.tech/tags/business.md>), [cloud-computing](<https://devfeed.tech/tags/cloud-computing.md>), [community](<https://devfeed.tech/tags/community.md>), [cybersecurity](<https://devfeed.tech/tags/cybersecurity.md>), [deal](<https://devfeed.tech/tags/deal.md>), [designline](<https://devfeed.tech/tags/designline.md>), [digital](<https://devfeed.tech/tags/digital.md>), [education](<https://devfeed.tech/tags/education.md>), [ee-times](<https://devfeed.tech/tags/ee-times.md>), [ee-times-university](<https://devfeed.tech/tags/ee-times-university.md>), [eelife](<https://devfeed.tech/tags/eelife.md>), [eelive-electronic-design](<https://devfeed.tech/tags/eelive-electronic-design.md>), [eet](<https://devfeed.tech/tags/eet.md>), [electrical-engineering](<https://devfeed.tech/tags/electrical-engineering.md>), [electronic-components](<https://devfeed.tech/tags/electronic-components.md>), [electronic-engineering-times](<https://devfeed.tech/tags/electronic-engineering-times.md>), [electronics](<https://devfeed.tech/tags/electronics.md>), [embedded-systems](<https://devfeed.tech/tags/embedded-systems.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [enterprise-ai](<https://devfeed.tech/tags/enterprise-ai.md>), [fundamentals-courses](<https://devfeed.tech/tags/fundamentals-courses.md>), [inference](<https://devfeed.tech/tags/inference.md>), [learning](<https://devfeed.tech/tags/learning.md>), [mergers-acquisitions](<https://devfeed.tech/tags/mergers-acquisitions.md>), [mixed-signal](<https://devfeed.tech/tags/mixed-signal.md>), [news](<https://devfeed.tech/tags/news.md>), [open-weight-ai](<https://devfeed.tech/tags/open-weight-ai.md>), [open-weight-ai-models](<https://devfeed.tech/tags/open-weight-ai-models.md>), [opinion](<https://devfeed.tech/tags/opinion.md>), [palo-alto-networks](<https://devfeed.tech/tags/palo-alto-networks.md>), [part-search](<https://devfeed.tech/tags/part-search.md>), [platform](<https://devfeed.tech/tags/platform.md>), [post-training](<https://devfeed.tech/tags/post-training.md>), [product-search](<https://devfeed.tech/tags/product-search.md>), [security](<https://devfeed.tech/tags/security.md>), [semiconductors](<https://devfeed.tech/tags/semiconductors.md>), [software](<https://devfeed.tech/tags/software.md>), [startup](<https://devfeed.tech/tags/startup.md>), [tech-papers](<https://devfeed.tech/tags/tech-papers.md>), [technology](<https://devfeed.tech/tags/technology.md>), [video](<https://devfeed.tech/tags/video.md>)

### AI overview

Palo Alto Networks plans to acquire Applied Compute for $4.5 billion in an all-stock deal, extending its strategy from cybersecurity into enterprise AI infrastructure. Applied Compute provides post-training and inference technology for open-weight models, enabling organizations to customize and run them. The article says this could let Palo Alto Networks embed security into model training, governance and deployment across public and sovereign clouds. It also describes risks tied to open-weight models, including tampering, data leakage and misuse; adapted weights may retain fragments of confidential training data. The company says customer-held keys support data sovereignty, but that capability does not yet extend to AI accelerators. The deal requires regulatory approval and other customary conditions, with closing expected in the third quarter of 2027.

### Source excerpt

Enterprise security outfit Palo Alto Networks snaps up a startup with a platform for post-training and inference of open-weight AI models. The post Palo Alto Networks Expands into Enterprise AI with $4.5B Deal appeared first on EE Times.

## Datadog LLM Observability: Features, Pricing, and Alternatives

DevFeed: [Datadog LLM Observability: Features, Pricing, and Alternatives](<https://devfeed.tech/articles/datadog-llm-observability-features-pricing-and-alternatives-145962.md>)

Original publisher: [Read original article](<https://newrelic.com/blog/ai/datadog-llm-observability>)

Author: John Blust

Published: 2026-11-20T11:51:52.599709Z

Content type: article

Language: en

Sources: [New Relic](<https://devfeed.tech/sources/new-relic.md>)

Topics: [LLM observability](<https://devfeed.tech/topics/llm-observability.md>), [datadog](<https://devfeed.tech/topics/datadog.md>), [agent observability](<https://devfeed.tech/topics/agent-observability.md>), [ai observability](<https://devfeed.tech/topics/ai-observability.md>), [observability](<https://devfeed.tech/topics/observability.md>), [tracing](<https://devfeed.tech/topics/tracing.md>), [OpenTelemetry](<https://devfeed.tech/topics/opentelemetry.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [phoenix](<https://devfeed.tech/topics/phoenix.md>), [LangChain](<https://devfeed.tech/topics/langchain.md>), [dynatrace](<https://devfeed.tech/topics/dynatrace.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-monitoring](<https://devfeed.tech/tags/ai-monitoring.md>), [alternatives](<https://devfeed.tech/tags/alternatives.md>), [compare](<https://devfeed.tech/tags/compare.md>), [datadog](<https://devfeed.tech/tags/datadog.md>), [devops](<https://devfeed.tech/tags/devops.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [features](<https://devfeed.tech/tags/features.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-observability](<https://devfeed.tech/tags/llm-observability.md>), [monitoring](<https://devfeed.tech/tags/monitoring.md>), [new-relic](<https://devfeed.tech/tags/new-relic.md>), [observability](<https://devfeed.tech/tags/observability.md>), [pricing](<https://devfeed.tech/tags/pricing.md>), [tracing](<https://devfeed.tech/tags/tracing.md>)

### AI overview

Datadog LLM Observability, now called Agent Observability, traces model calls and agent workflows, evaluates quality and safety, supports prompt experiments, and tracks latency, tokens, errors, and costs. It is a standalone product billed by model-call spans; tool, retrieval, and workflow spans do not count. Both listed plans include core tracing and evaluation features, and annual Pro pricing starts at $160 per month, with retention add-ons affecting per-span rates. Alternatives differ in approach: New Relic connects AI telemetry to existing APM, including optional eBPF monitoring; Langfuse offers a self-hostable core; LangSmith emphasizes agent tracing and evaluation workflows; and Arize Phoenix supports local deployment with OpenTelemetry-based instrumentation. The text notes that Datadog can correlate with its other products through shared trace IDs and supports Python, Node.js, Java, OpenTelemetry, and HTTP API integration.

### Source excerpt

An objective look at Datadog LLM Observability: tracing, evaluations, how pricing works, and how alternatives like New Relic AI Monitoring compare.

## Agentic AI repeats same old enterprise adoption cycle

DevFeed: [Agentic AI repeats same old enterprise adoption cycle](<https://devfeed.tech/articles/agentic-ai-repeats-same-old-enterprise-adoption-cycle-146087.md>)

Original publisher: [Read original article](<https://www.infoworld.com/article/4249734/agentic-ai-repeats-same-old-enterprise-adoption-cycle.html>)

Author: David Linthicum

Published: 2026-11-20T09:00:00Z

Content type: opinion

Language: en

Sources: [InfoWorld](<https://devfeed.tech/sources/infoworld.md>)

Topics: [Agentic AI](<https://devfeed.tech/topics/what-is-agentic-ai.md>), [Enterprise Architecture](<https://devfeed.tech/topics/enterprise-architecture.md>), [AI Integration Strategies](<https://devfeed.tech/topics/ai-integration-strategies.md>), [agentic workflows](<https://devfeed.tech/topics/agentic-workflows.md>), [business logic](<https://devfeed.tech/topics/business-logic.md>), [API Governance](<https://devfeed.tech/topics/api-governance.md>), [observability](<https://devfeed.tech/topics/observability.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>)

Tags: [adoption](<https://devfeed.tech/tags/adoption.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-ai](<https://devfeed.tech/tags/agentic-ai.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [apis](<https://devfeed.tech/tags/apis.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [artificial-intelligence](<https://devfeed.tech/tags/artificial-intelligence.md>), [artificial-intelligence-emerging-technology-generative-ai](<https://devfeed.tech/tags/artificial-intelligence-emerging-technology-generative-ai.md>), [automate](<https://devfeed.tech/tags/automate.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [cost](<https://devfeed.tech/tags/cost.md>), [costs](<https://devfeed.tech/tags/costs.md>), [data](<https://devfeed.tech/tags/data.md>), [decision](<https://devfeed.tech/tags/decision.md>), [emerging-technology](<https://devfeed.tech/tags/emerging-technology.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [exception-handling](<https://devfeed.tech/tags/exception-handling.md>), [generative-ai](<https://devfeed.tech/tags/generative-ai.md>), [guardrails](<https://devfeed.tech/tags/guardrails.md>), [language-en](<https://devfeed.tech/tags/language-en.md>), [logging](<https://devfeed.tech/tags/logging.md>), [observability](<https://devfeed.tech/tags/observability.md>), [production](<https://devfeed.tech/tags/production.md>), [project](<https://devfeed.tech/tags/project.md>), [storytype-opinion](<https://devfeed.tech/tags/storytype-opinion.md>)

### AI overview

A distributor's agentic AI sales system ran successfully, but after supporting integration, governance, security, testing and operational costs were included, it cost roughly 20 times as much as a traditional architecture for about the same ROI. The article argues that agents are most useful when work requires interpretation, adaptation and semi-autonomous action; stable, rules-based or highly regulated workflows may be served more economically by conventional applications, workflow engines, rules or API integration. It recommends starting with measurable business value and assessing production costs and governance needs rather than adopting an architecture because it is fashionable. The text compares this adoption pattern with SOA, microservices and cloud-native systems, each of which can deliver benefits but also impose complexity when used without a fitting need or organizational maturity.

### Source excerpt

The pitch sounded perfect. A large regional distributor needed to build a new sales transaction management system. The goal was straightforward: automate parts of order intake, discount approvals, exception handling, customer communications, and inventory checks. It was not a trivial system, but it was also something the industry had built before. Then the vendor presentations began. Every software provider seemed to have a new agentic AI story. Their platforms had agents that could make decisions, orchestrate workflows, interpret customer intent, route exceptions, and learn from business context. One vendor described its offering as a common LLM brain that could make decisions as if the company had hired 10,000 experts, except those experts were not on the payroll. That line was from an actual sales pitch, and it landed well in the boardroom. The company decided to go agentic. The architecture would use AI agents to manage decision flows, invoke tools, check policies, interact with enterprise systems, and escalate only when needed. It looked modern. It looked strategic. It looked like the future. Then reality arrived. Building the system required far more integration work than expected. Each agent needed access to clean data, governed APIs, business rules, security controls, observability, exception handling, and human override mechanisms. The team had to build guardrails around decisions, logging around reasoning, test harnesses for unpredictable outcomes, and new operational processes for things traditional systems handled more predictably. Costs ballooned. Security reviews took longer. Operational complexity increased. The project ultimately ran fine once in production, but six months later, the company realized the uncomfortable truth: the system cost roughly 20 times more than a traditional architecture would have cost when all supporting complexity was included, and it delivered about the same ROI. I am hearing versions of this story once or twice a week from

## Attacks and Defenses in Decentralized Multi-Agent Systems with Hop-by-Hop Task Decomposition and Delegation

DevFeed: [Attacks and Defenses in Decentralized Multi-Agent Systems with Hop-by-Hop Task Decomposition and Delegation](<https://devfeed.tech/articles/attacks-and-defenses-in-decentralized-multi-agent-systems-with-hop-by-hop-task-decomposition-and-delegation-146019.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25056>)

Author: Matthew D. Laws, Maria Lazou, Alina Oprea, Cristina Nita-Rotaru

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: [ai security](<https://devfeed.tech/topics/ai-security.md>), [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [hierarchical agent systems](<https://devfeed.tech/topics/hierarchical-agent-systems.md>), [Cybersecurity](<https://devfeed.tech/topics/cybersecurity.md>), [agent to agent communication](<https://devfeed.tech/topics/agent-to-agent-communication.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [availability](<https://devfeed.tech/tags/availability.md>), [cryptographic](<https://devfeed.tech/tags/cryptographic.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [decentralized](<https://devfeed.tech/tags/decentralized.md>), [delegation](<https://devfeed.tech/tags/delegation.md>), [execution](<https://devfeed.tech/tags/execution.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [latency](<https://devfeed.tech/tags/latency.md>), [llm](<https://devfeed.tech/tags/llm.md>), [malicious](<https://devfeed.tech/tags/malicious.md>), [memory](<https://devfeed.tech/tags/memory.md>), [mitigation](<https://devfeed.tech/tags/mitigation.md>), [model](<https://devfeed.tech/tags/model.md>), [multi-agent](<https://devfeed.tech/tags/multi-agent.md>), [multi-agent-systems](<https://devfeed.tech/tags/multi-agent-systems.md>), [overhead](<https://devfeed.tech/tags/overhead.md>), [protocol](<https://devfeed.tech/tags/protocol.md>), [runtime](<https://devfeed.tech/tags/runtime.md>), [security](<https://devfeed.tech/tags/security.md>)

### AI overview

The abstract reports that one malicious agent in a decentralized multi-agent LLM system can cause up to 100% task failure, raise token use by up to 7.7x and runtime by up to 3.2x, or exfiltrate up to 94% of relevant prior tasks. The proposed cryptographic protocol creates a verifiable execution trace, which an auditor analyzes to detect and attribute malicious behavior. On HiddenBench, evaluations with gpt-6.1-sol and Qwen3.8-27B-FP8 found negligible latency overhead and 98% detection and correct attribution of attacks, including attacks that complied with the protocol.

### Source excerpt

arXiv:2611.25056v1 Announce Type: new Abstract: Decentralized multi-agent LLM systems enable independently administered agents, each with its own operator, model, memory, and configuration, to collaborate on complex tasks without centralized coordination. Agents autonomously decompose tasks, execute subtasks and delegate remaining work to peers, enabling scalability and eliminating single points of failure. However, decentralization also introduces significant security challenges, as no single entity has complete visibility into task execution or the authority to enforce correct behavior across agents. In this paper, we systematize threats from malicious agents in decentralized multi-agent systems and propose mitigation techniques. We first design and evaluate attacks against integrity, availability, and confidentiality, demonstrating that a single malicious agent can cause up to 100% task failure, inflate token consumption and runtime by up to 7.7x and 3.2x, respectively, and exfiltrate up to 94% of relevant prior tasks from benign agents. We then present a cryptographic protocol that produces an irrefutable verified execution trace of every action taken in the system and develop an auditor that analyzes these traces to detect and attribute malicious behavior. Evaluations on HiddenBench using gpt-6.1-sol and Qwen3.8-27B-FP8 show that our protocol incurs negligible latency overhead, while the auditor detects and correctly attributes 98% of attacks, including those that comply with the protocol.

## Hardening JavaScript Programs against LLM-Based Vulnerability Detection

DevFeed: [Hardening JavaScript Programs against LLM-Based Vulnerability Detection](<https://devfeed.tech/articles/hardening-javascript-programs-against-llm-based-vulnerability-detection-146045.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25718>)

Author: Zhenhao Tian, Jintong Yu, Jingwei Feng, Jianwei Zhuge, Penghui Li

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: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Compilers and Linkers](<https://devfeed.tech/topics/compilers-and-linkers.md>), [Fuzzing/Fuzz testing](<https://devfeed.tech/topics/fuzzing.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense](<https://devfeed.tech/tags/defense.md>), [hardening](<https://devfeed.tech/tags/hardening.md>), [javascript](<https://devfeed.tech/tags/javascript.md>), [llm](<https://devfeed.tech/tags/llm.md>), [vulnerability](<https://devfeed.tech/tags/vulnerability.md>), [vulnerability-detection](<https://devfeed.tech/tags/vulnerability-detection.md>)

### AI overview

AntiAgent hardens programs before release by applying semantics-preserving transformations designed to impede LLM-based vulnerability detection. It evolves programs with known vulnerabilities, tracks when a detection agent misses or misreports them, and turns successful transformations into reusable rules. For a new program, an LLM and those rules work together to adapt the transformations. In tests on real-world JavaScript applications, AntiAgent reduced vulnerabilities detected by each of four agents, including Claude Code and Codex, by 52.0-60.7%, outperforming JS-Obfuscator and FPA on every agent. Transformed programs passed all project tests. The abstract also reports effectiveness on vulnerability types absent from the training corpus and on Python programs.

### Source excerpt

arXiv:2611.25718v1 Announce Type: new Abstract: JavaScript programs are often distributed as source code, so attackers can analyze them for vulnerabilities. LLM-based vulnerability detection agents now make this analysis much easier. To mitigate this emerging threat, we present AntiAgent, a defense that lets developers harden their programs against such detection agents. AntiAgent transforms a program before release to impede an agent's analysis. To find effective transformations, AntiAgent evolves a corpus of programs with known vulnerabilities and traces how a detection agent analyzes each evolved version. It keeps the versions on which the agent misses or misreports the vulnerability and extracts their transformations into reusable rules. To protect a new program, AntiAgent carefully divides the task between an LLM and the rules, so that the transformations adapt to the program while preserving its semantics. We evaluate AntiAgent on real-world JavaScript applications against four detection agents, including Claude Code and Codex. AntiAgent reduces the number of vulnerabilities that each agent detects by 52.0-60.7%, outperforming the conventional obfuscator JS-Obfuscator and the LLM-based transformation method FPA on every agent. The transformed programs pass all project tests, indicating that AntiAgent preserves the original program semantics. The transformation rules also remain effective on vulnerability types absent from the corpus and on programs written in another language like Python.

## Why Trust Your Agent? Empirical Security Gains from TRiSM-Guided Agentic Workflows in Healthcare

DevFeed: [Why Trust Your Agent? Empirical Security Gains from TRiSM-Guided Agentic Workflows in Healthcare](<https://devfeed.tech/articles/why-trust-your-agent-empirical-security-gains-from-trism-guided-agentic-workflows-in-healthcare-145354.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.28666>)

Author: Liam Kearns

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

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>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [least privilege](<https://devfeed.tech/topics/least-privilege.md>), [Security](<https://devfeed.tech/topics/security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-workflows](<https://devfeed.tech/tags/agentic-workflows.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-trust](<https://devfeed.tech/tags/ai-trust.md>), [attack](<https://devfeed.tech/tags/attack.md>), [automation](<https://devfeed.tech/tags/automation.md>), [breaches](<https://devfeed.tech/tags/breaches.md>), [claude-haiku-4-5](<https://devfeed.tech/tags/claude-haiku-4-5.md>), [client](<https://devfeed.tech/tags/client.md>), [construction](<https://devfeed.tech/tags/construction.md>), [covid-19](<https://devfeed.tech/tags/covid-19.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [data](<https://devfeed.tech/tags/data.md>), [diagnosis](<https://devfeed.tech/tags/diagnosis.md>), [disclosure](<https://devfeed.tech/tags/disclosure.md>), [gemini](<https://devfeed.tech/tags/gemini.md>), [generation](<https://devfeed.tech/tags/generation.md>), [healthcare](<https://devfeed.tech/tags/healthcare.md>), [security](<https://devfeed.tech/tags/security.md>), [trust](<https://devfeed.tech/tags/trust.md>)

### AI overview

In a medical report-generation application tested with five LLMs, a TRiSM-guided workflow reduced attack success for RAG poisoning (31% to 10%) and data-field injection (42% to 25%), eliminated client-side network injection by construction, and increased report accuracy from 72.5% to 86.5%. In an exploratory grouping, attacks aimed beyond the reporting task, including data disclosure and tool abuse, fell from 21.7% to 2.5%, with no patient records disclosed under the redesign. However, attacks that changed the Covid-19 diagnosis remained effective, making up 32 of 35 successful attacks against the redesign. The findings suggest least privilege limits what injections can reach but cannot ensure the model resists injected content within its task; clinical outputs still need independent verification. No tested model resisted all three attack vectors.

### Source excerpt

arXiv:2606.28666v2 Announce Type: replace-cross Abstract: Agent-based AI has enabled the automation of tasks by exposing application tools and resources to large language models (LLMs). However, to improve scope and accuracy, agents are often given access rights that exceed those of ordinary users, introducing significant security risks and potential regulatory breaches. This paper applies the AI Trust, Risk, and Security Management (TRiSM) framework to a medical report-generation application to demonstrate how an insecure baseline workflow can be transformed into a security-conscious agentic workflow. Both designs were evaluated with five LLMs (Claude Haiku 4.5, GPT-4.1-nano, GPT-4.1-mini, GPT-5.4-mini and Gemini 2.5 Flash) across 800 report generations and 500 attack trials covering RAG poisoning, data-field injection and client-side network injection. The redesign reduced attack success from 31% to 10% for RAG poisoning and from 42% to 25% for data-field injection (McNemar p < 0.001 for both), removed network injection by construction, and raised report accuracy from 72.5% to 86.5%. In an exploratory post-hoc grouping by attack goal, success against attacks aimed outside the report task, such as data disclosure and tool abuse, fell from 21.7% to 2.5% and no patient record was disclosed under the redesign. Attacks altering the reported Covid-19 diagnosis remained effective, accounting for 32 of the 35 successful attacks against the redesign. Least-privilege architecture limits what an injection can reach, but not whether a model follows injected content within its task, so clinical outputs need verification independent of the LLM. No model resisted all three attack vectors, so model selection cannot substitute for secure workflow design.

## Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)

DevFeed: [Replication of Agentic Security Events: Audit and Investigation (ROGUE-AI)](<https://devfeed.tech/articles/replication-of-agentic-security-events-audit-and-investigation-rogue-ai-145684.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26553>)

Author: Mengxiao Wang, Mahbub Alam, Kalyan Nakka, Nitesh Saxena

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

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>), [audit](<https://devfeed.tech/topics/audit.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [Security](<https://devfeed.tech/topics/security.md>)

Tags: [agent-framework](<https://devfeed.tech/tags/agent-framework.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agentic-security](<https://devfeed.tech/tags/agentic-security.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [audit](<https://devfeed.tech/tags/audit.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [database](<https://devfeed.tech/tags/database.md>), [deepseek](<https://devfeed.tech/tags/deepseek.md>), [delete](<https://devfeed.tech/tags/delete.md>), [events](<https://devfeed.tech/tags/events.md>), [failures](<https://devfeed.tech/tags/failures.md>), [files](<https://devfeed.tech/tags/files.md>), [framework](<https://devfeed.tech/tags/framework.md>), [gemini](<https://devfeed.tech/tags/gemini.md>), [investigation](<https://devfeed.tech/tags/investigation.md>), [multi-agent](<https://devfeed.tech/tags/multi-agent.md>), [replication](<https://devfeed.tech/tags/replication.md>), [rogue-ai](<https://devfeed.tech/tags/rogue-ai.md>), [security](<https://devfeed.tech/tags/security.md>), [security-events](<https://devfeed.tech/tags/security-events.md>)

### AI overview

ROGUE-AI audits reported agent failures by reconstructing incidents offline and comparing agent runs with matched controls that omit only the suspected cause. In 108 attempted replications, it reproduced reported harm in 89%; across five GPT, Claude, Gemini, and DeepSeek models, the reported mistake occurred in 18-33% of runs versus at most 1% in controls. Among 72 incidents involving real-world harm, 62 had no attacker: agents followed their tasks but missed ordinary hazards. These replications caused more harm than prompt-injection cases for all five models, exposing a gap for defenses focused on injected instructions. A warning after each tool output that named the hazard and its consequence reduced harm for all five models; an equally long generic note left most harm in place. The abstract presents the replications as a reusable audit framework vendors can rerun before releases.

### Source excerpt

arXiv:2611.26553v1 Announce Type: cross Abstract: Agents based on large language models (LLMs) can now delete files and send payments on behalf of their users. News media reports routinely describe failures of these agents (e.g., Replit's coding agent deleting a production database), but a report often only says what went wrong and rarely (in)validates the incident or finds its root cause. Incident catalogs collect these reports, and benchmarks record whether agents fail, but neither replicates an incident nor tests the cause its report might identify. In this paper, we present ROGUE-AI (Replication of Agentic Security Events: Audit and Investigation), a multi-agent framework that closes the above gap and rebuilds each reported incident offline together with a matched control that lacks only the blamed cause, and runs agents on both. Our results suggest that ROGUE-AI is able to reproduce the reported harm in 89% of its 108 attempted replications. Agents built on five current GPT, Claude, Gemini and DeepSeek models made the reported mistake in 18% to 33% of runs, against at most 1% without the blamed cause. Of the 72 incidents with real-world harm, 62 had no underlying attacker, and in these, agents did what their task asked for but missed an ordinary hazard, such as a database shared with the live site. Replications without an attacker caused more harm than replications of prompt injection for all five models. Defenses against prompt injection look for injected instructions, but these replications contain none. Further, a warning after each tool output that names the hazard and its consequence lowered harm for all five models, whereas a note of the same length that does not mention the hazard left most of the harm in place. Built from real incidents, the ROGUE-AI replications form a reusable audit framework that vendors can rerun before each new release.

## What the model knows but does not say: monitoring medical LLMs through verbalized thoughts

DevFeed: [What the model knows but does not say: monitoring medical LLMs through verbalized thoughts](<https://devfeed.tech/articles/what-the-model-knows-but-does-not-say-monitoring-medical-llms-through-verbalized-thoughts-145455.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24994>)

Author: Frederik Hauke, Jan Clusmann, Hanna Kreutzer, Christiane Kuhl, Tianyu Han, Felix Barajas Ordonez, Sven Nebelung, Jakob Nikolas Kather, Daniel Truhn

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

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

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

Tags: [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [gemma-3](<https://devfeed.tech/tags/gemma-3.md>), [llama-3-3](<https://devfeed.tech/tags/llama-3-3.md>), [llms](<https://devfeed.tech/tags/llms.md>), [medical](<https://devfeed.tech/tags/medical.md>), [monitoring](<https://devfeed.tech/tags/monitoring.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [thoughts](<https://devfeed.tech/tags/thoughts.md>)

### AI overview

The study reports that verbalized internal activations can reveal clinical prompt injections that ordinary answers conceal. A natural-language autoencoder converts Gemma-3-27B activations into text; an input-blind judge assesses that "thoughts" text and the model's answer separately. Across two attack objectives (omitting pathology or covertly recommending a drug), two input modalities, and two delivery mechanisms, the thoughts channel detected injections significantly more often in all eight conditions (p<0.05), while both channels had near-zero false positives on clean inputs. The omission finding also carried over to Llama-3.3-70B. The channel flagged unsafe requests while the model was still reading them. The approach used an off-the-shelf verbalizer and general-purpose judge, without retraining; the excerpt does not establish performance beyond these reported settings.

### Source excerpt

arXiv:2611.24994v1 Announce Type: cross Abstract: Efforts to keep clinical language models safe have concentrated on what enters and leaves the model, but a manipulated model can return an answer that looks entirely ordinary. Here we show that the model's own internal state, rendered in plain language, exposes manipulations the output conceals. Using a natural-language autoencoder, we translate the activations of a multimodal LLM (Gemma-3-27B) into text and let an input-blind judge read this verbalized internal state (the thoughts channel) and the model's answer separately. Across two clinical attack objectives (omitting a pathology, covertly recommending a drug), two input modalities, and two delivery mechanisms, the thoughts channel caught the injection significantly more often than the answer in all eight injected conditions (p<0.05), while both channels held near-zero false-positive rates on clean inputs; the omission result carried over to Llama-3.3-70B. The thoughts channel also registered unsafe clinical requests as harmful while the model was still reading them, before it decided whether to refuse. None of this requires retraining: an off-the-shelf verbalizer and a general-purpose judge are enough.

## HireWatch: Evaluating LLM Compliance with U.S. Employment Law Under Contextual Pressure

DevFeed: [HireWatch: Evaluating LLM Compliance with U.S. Employment Law Under Contextual Pressure](<https://devfeed.tech/articles/hirewatch-evaluating-llm-compliance-with-u-s-employment-law-under-contextual-pressure-145449.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24970>)

Author: Huy Nghiem, Maria Kostylew, Noam Kolt

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

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

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

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cy](<https://devfeed.tech/tags/cs-cy.md>), [employment-law](<https://devfeed.tech/tags/employment-law.md>), [llm-compliance](<https://devfeed.tech/tags/llm-compliance.md>)

### AI overview

HireWatch tests whether LLMs select interview questions that comply with U.S. employment law. It uses 200 questions classified under federal statutes, regulations, and agency guidance, and measures unlawful selections across occupations and contexts. In about 100,000 trials with 18 closed and open-weight models, unlawful selections rose sharply under greater contextual pressure. Explicit instructions to comply improved results but did not prevent all violations. The evidence provided is an abstract, so it does not give the specific pressure conditions or model-by-model results.

### Source excerpt

arXiv:2611.24970v1 Announce Type: cross Abstract: The deployment of large language models (LLMs) in hiring pipelines is growing, yet existing evaluations generally overlook their compliance with legal rules. To fill this gap, we introduce HireWatch, an evaluation that assesses whether LLMs comply with U.S. laws that govern the questions that job candidates may be asked. We construct a 200-item bank of lawful and unlawful interview questions, whose legal classification is based on U.S. federal statutes, regulations, and agency guidance. We instruct LLMs to select questions from this bank in preparing for job interviews, across a set of various occupations and contexts, and report the fraction of unlawful questions that LLMs select. Evaluation of 18 closed and open-weight LLMs across ~100,000 trials under different types of contextual pressure reveals that LLMs' legal compliance is brittle. While LLMs generally avoid selecting unlawful questions in low-pressure contexts, legal violations increase sharply under higher pressure. Explicit prompts to comply with law can improve compliance rates, but do not altogether eliminate legal violations. We publicly release our code and data at https://github.com/hnghiem-nlp/hirewatch_repo.

## Living off the Local Model: Abusing Browser-Embedded Language Models for Post-Exploitation

DevFeed: [Living off the Local Model: Abusing Browser-Embedded Language Models for Post-Exploitation](<https://devfeed.tech/articles/living-off-the-local-model-abusing-browser-embedded-language-models-for-post-exploitation-145464.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25086>)

Author: William Fleshman, Joseph Lucas

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

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>), [Local AI](<https://devfeed.tech/topics/local-ai.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Google Chrome](<https://devfeed.tech/topics/google-chrome.md>), [Cryptography](<https://devfeed.tech/topics/cryptography.md>), [function calling](<https://devfeed.tech/topics/function-calling.md>)

Tags: [browser](<https://devfeed.tech/tags/browser.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>), [embedded](<https://devfeed.tech/tags/embedded.md>), [exploitation](<https://devfeed.tech/tags/exploitation.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [local](<https://devfeed.tech/tags/local.md>), [post-exploitation](<https://devfeed.tech/tags/post-exploitation.md>)

### AI overview

The paper reports that Chrome's on-device Gemini Nano can be driven from the host using operating-system interpreters, and that its one-time user-activation check for downloading the model can be met with a simulated click or bypassed. The authors test credential identification, PII detection in email, image classification, and custom tool use. In a simulated operation, they combine these capabilities to read a password from a webcam image, unlock and triage an encrypted archive, and stage a private key for exfiltration without user interaction. These findings describe the current model and a simulated scenario.

### Source excerpt

arXiv:2611.25086v1 Announce Type: cross Abstract: Browsers now ship local language models to ordinary endpoints. Google includes Gemini Nano in Chrome to enable productivity features such as summarization, with inference running entirely on the device. These browser-embedded models provide new living-off-the-land opportunities for attackers, each instance a potential post-exploitation engine already installed, signed by the vendor, and invisible to network monitoring. The Gemini model is multimodal, capable of processing files, analyzing screenshots, or conversing with an attacker. We show that it can be driven from the host using only interpreters that ship with the operating system and find that the one-time user-activation check before the model is downloaded is easily satisfied with a simulated click or bypassed entirely.We measure and report the current model's ability to identify credentials on the host, detect PII in email, classify images, and call custom tools. We combine these capabilities and use the model to carry out a simulated cyber operation: reading a password from a webcam image, unlocking an encrypted archive, triaging the contents, and staging a private key for exfiltration, end-to-end without user-interaction.

## Why Do LLMs Help with Scientific Misconduct Requests and What to Do About It?: Increasing the Trustworthiness of AI for Science

DevFeed: [Why Do LLMs Help with Scientific Misconduct Requests and What to Do About It?: Increasing the Trustworthiness of AI for Science](<https://devfeed.tech/articles/why-do-llms-help-with-scientific-misconduct-requests-and-what-to-do-about-it-increasing-the-trustworthiness-of-ai-for-science-145523.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25492>)

Author: Almene De Meran Meguimtsop, Daniel E. Acuna

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

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [AI for science](<https://devfeed.tech/topics/ai-for-science.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [llms](<https://devfeed.tech/tags/llms.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [prompts](<https://devfeed.tech/tags/prompts.md>), [research-integrity](<https://devfeed.tech/tags/research-integrity.md>), [safety](<https://devfeed.tech/tags/safety.md>), [science](<https://devfeed.tech/tags/science.md>)

### AI overview

Across eight LLMs tested on 1,350 SciIntBench prompts, models often recognized scientific misconduct requests but rarely refused them: they identified misconduct in 91% of cases when asked directly, yet refused in 34%. Request framing was associated with refusal; changing instructions into questions about permissibility raised refusal from 52% to 88%. Two tested system-prompt changes--asking models to assess a request before answering and reminding them of responsible conduct of research principles--raised refusal on adversarial requests from 58% to 87% and 86%, respectively. The abstract reports these as study findings, not evidence that the changes prevent misconduct in practice.

### Source excerpt

arXiv:2611.25492v1 Announce Type: new Abstract: Large language models (LLMs) support scientific research, but some of that assistance can also undermine research integrity. Existing safety evaluations in science often measure whether models refuse misconduct requests, but rarely explain whether they even recognize the problem or why they comply. We investigate this gap across 8 LLMs using SciIntBench, an existing benchmark of 1,350 prompts about scientific research integrity. We examine the reasons behind compliance in four steps. First, when asked directly whether fulfilling a request would involve misconduct, models overwhelmingly identify such requests, even in the Most Covert cases (91%), while they refuse to carry out the request in only 34% of cases. Second, we examine which request characteristics are associated with refusal using features that capture how the request is framed, how much the assistant has been told about what it is helping to hide, and how the goal and the user's stake are stated, among others. Third, we estimate that the most significant effect was to rephrase requests as questions about permissibility rather than instructions, which increased refusal from 52% to 88%. Finally, we test two system-prompt fixes: asking models to assess the request before responding and reminding them of responsible conduct of research principles. These changes increase refusal rates on adversarial requests from 58% to 87% and 86%, respectively. Our findings show that models recognize wrongful requests but still help users violate integrity. Understanding why these failures occur and testing ways to reduce them offer promising paths toward more trustworthy AI for science.

## ALIoTh: Orchestrating Multi-Agent LLMs for Autonomous IoT Penetration Testing

DevFeed: [ALIoTh: Orchestrating Multi-Agent LLMs for Autonomous IoT Penetration Testing](<https://devfeed.tech/articles/alioth-orchestrating-multi-agent-llms-for-autonomous-iot-penetration-testing-146047.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25809>)

Author: Pironti Francesco Aurelio, Arena Luigi, Blefari Francesco, Lupinacci Matteo, Romeo Francesco, Maresca Alessio

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: [Hacking](<https://devfeed.tech/topics/hacking.md>), [Internet of things](<https://devfeed.tech/topics/iot.md>), [Multi Agent Systems](<https://devfeed.tech/topics/multi-agent-systems.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Enhancing AI model performance](<https://devfeed.tech/topics/enhancing-ai-model-performance.md>), [smartdevices](<https://devfeed.tech/topics/smartdevices.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [assessment](<https://devfeed.tech/tags/assessment.md>), [attack](<https://devfeed.tech/tags/attack.md>), [autonomous](<https://devfeed.tech/tags/autonomous.md>), [cameras](<https://devfeed.tech/tags/cameras.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [deployment](<https://devfeed.tech/tags/deployment.md>), [devices](<https://devfeed.tech/tags/devices.md>), [enumeration](<https://devfeed.tech/tags/enumeration.md>), [exploit](<https://devfeed.tech/tags/exploit.md>), [exploitation](<https://devfeed.tech/tags/exploitation.md>), [framework](<https://devfeed.tech/tags/framework.md>), [gathering](<https://devfeed.tech/tags/gathering.md>), [generation](<https://devfeed.tech/tags/generation.md>), [group-policy](<https://devfeed.tech/tags/group-policy.md>), [iiot](<https://devfeed.tech/tags/iiot.md>), [iot](<https://devfeed.tech/tags/iot.md>), [llms](<https://devfeed.tech/tags/llms.md>), [multi-agent](<https://devfeed.tech/tags/multi-agent.md>), [penetration-testing](<https://devfeed.tech/tags/penetration-testing.md>)

### AI overview

ALIoTh is a workflow-driven framework that coordinates specialized LLM agents through LangGraph and a Self-RAG architecture to automate IoT and IIoT penetration testing. Its described pipeline covers network, device, and service enumeration, vulnerability assessment, payload generation, reverse-shell deployment, and privilege escalation. Self-RAG gives agents contextual access to IoT security documentation, tool manuals, and known attack techniques, supporting scenarios involving multiple devices and group-policy manipulation. The supplied text describes the approach but provides no evaluation results or comparative performance.

### Source excerpt

arXiv:2611.25809v1 Announce Type: new Abstract: Penetration testing in Internet of Things (IoT) environments, especially those composed of various different devices, e.g., smart plugs, smart cameras, and smart doors, presents great challenges across the phases of information gathering, attack surface enumeration, privilege escalation, and lateral movement. This paper introduces ALIoTh, a multi-agent, workflow-driven penetration testing framework that integrates Large Language Models agents within a Self-Retrieval-Augmented Generation (Self-RAG) architecture to autonomously identify and exploit vulnerabilities in IoT and Industrial-IoT (IIoT) environments. By leveraging state-of-the-art language models and orchestrating specialized agents via the LangGraph framework, ALIoTh dynamically plans and executes a full-spectrum offensive pipeline. This includes network, devices, and service enumeration, vulnerability assessment, payload generation, reverse shell deployment, and privilege escalation on each device. The Self-RAG mechanism equips expert agents with contextual access to IoT security knowledge, e.g., related documentation, tool manuals, and known attack techniques, enhancing their reasoning capabilities in scenarios involving multiple device exploitation and group policy manipulation.

## Model Compression vs. Adversarial Robustness: An Empirical Study on Language Models for Code

DevFeed: [Model Compression vs. Adversarial Robustness: An Empirical Study on Language Models for Code](<https://devfeed.tech/articles/model-compression-vs-adversarial-robustness-an-empirical-study-on-language-models-for-code-145758.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2508.03949>)

Author: Md. Abdul Awal, Mrigank Rochan, Chanchal K. Roy

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

Content type: article

Language: en

Sources: [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [Post-training optimization](<https://devfeed.tech/topics/post-training-optimization.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: [adversarial](<https://devfeed.tech/tags/adversarial.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [code](<https://devfeed.tech/tags/code.md>), [compression](<https://devfeed.tech/tags/compression.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [impact](<https://devfeed.tech/tags/impact.md>), [inference](<https://devfeed.tech/tags/inference.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [performance](<https://devfeed.tech/tags/performance.md>), [pruning](<https://devfeed.tech/tags/pruning.md>), [quantization](<https://devfeed.tech/tags/quantization.md>), [real-world](<https://devfeed.tech/tags/real-world.md>), [robustness](<https://devfeed.tech/tags/robustness.md>), [security](<https://devfeed.tech/tags/security.md>), [size](<https://devfeed.tech/tags/size.md>), [software](<https://devfeed.tech/tags/software.md>), [speeds](<https://devfeed.tech/tags/speeds.md>), [strategies](<https://devfeed.tech/tags/strategies.md>)

### AI overview

The study finds that compression can preserve language models' usual performance on software analytics tasks while significantly reducing their robustness to adversarial attacks. It evaluates pruning, quantization, and knowledge distillation across compressed versions of three code language models, three tasks, six metrics, and four classical attacks. The results point to a trade-off between reducing model size and adversarial robustness, a concern for security-critical deployments. The abstract does not specify which compression methods or attacks caused the largest effects.

### Source excerpt

arXiv:2508.03949v3 Announce Type: replace Abstract: Transformer-based language models for code have shown remarkable performance in various software analytics tasks, but their adoption is hindered by high computational costs, slow inference speeds, and substantial environmental impact. Model compression techniques such as pruning, quantization, and knowledge distillation have gained traction in addressing these challenges. However, the impact of these strategies on the robustness of compressed language models for code in adversarial scenarios remains poorly understood. Understanding how these compressed models behave under adversarial attacks is essential for their safe and effective deployment in real-world applications. To bridge this knowledge gap, we conduct a comprehensive evaluation of how common compression strategies affect the adversarial robustness of compressed models. We assess the robustness of compressed versions of three widely used language models for code across three software analytics tasks, using six evaluation metrics and four commonly used classical adversarial attacks. Our findings indicate that compressed models generally maintain comparable performance to their uncompressed counterparts. However, when subjected to adversarial attacks, compressed models exhibit significantly reduced robustness. These results reveal a trade-off between model size reduction and adversarial robustness, underscoring the need for careful consideration when deploying compressed models in security-critical software applications. Our study highlights the need for further research into compression strategies that strike a balance between computational efficiency and adversarial robustness, which is essential for deploying reliable language models for code in real-world software applications.

## Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge

DevFeed: [Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge](<https://devfeed.tech/articles/safeguarding-llms-via-model-agnostic-latent-safety-signals-from-dark-knowledge-145395.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2610.07532>)

Author: Wonjun Lee, Kyungsik Yang, Gaeun Ji, Vaidehi Patil, Haon Park, Bumsub Ham, Mohit Bansal, Suhyun Kim

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

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>), [Language models](<https://devfeed.tech/topics/language-models.md>)

Tags: [attack](<https://devfeed.tech/tags/attack.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-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [defense](<https://devfeed.tech/tags/defense.md>), [discrimination](<https://devfeed.tech/tags/discrimination.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [hidden-states](<https://devfeed.tech/tags/hidden-states.md>), [llms](<https://devfeed.tech/tags/llms.md>), [model-agnostic](<https://devfeed.tech/tags/model-agnostic.md>), [robust](<https://devfeed.tech/tags/robust.md>), [safety](<https://devfeed.tech/tags/safety.md>), [tokenizers](<https://devfeed.tech/tags/tokenizers.md>), [utility](<https://devfeed.tech/tags/utility.md>)

### AI overview

LADE detects potentially harmful queries using safety signals in an LLM's first-token probability distribution beyond its most likely token. It selects tokens whose probabilities differ sharply between harmful and benign queries, maps them across tokenizers, then applies k-nearest-neighbor classification. The authors report that these signals align across diverse LLMs, enabling model-agnostic use without relying on architecture-specific hidden states. Across the evaluated models and benchmarks, LADE reportedly lowers jailbreak attack success rates while maintaining a competitive safety-utility trade-off. The supplied text gives no specific metrics or benchmark names.

### Source excerpt

arXiv:2610.07532v2 Announce Type: replace-cross Abstract: LLMs have advanced rapidly, raising growing concerns about their safety. Recent work has proposed approaches to detect and defend against attacks including defenses at decoding stage that leverage models' hidden states. However, existing decoding-stage defenses suffer from two limitations. First, they introduce a trade-off between safety and over-refusal, where strengthening safety degrades the model's helpfulness on benign queries. Second, many of these methods rely on internal hidden states and are thus restricted to specific architectures, incurring substantial overhead and limited generalization across models. To address these limitations, we introduce LADE (Latent Safety Signals for Defense), which leverages latent safety signals extracted by contrasting harmful and benign queries from dark knowledge (i.e., information carried by the output probability distribution beyond its argmax) in the first-token output probability distribution. Our key insight is that, beyond surface-level refusal tokens, the dark knowledge in the first-token distribution contains latent safety signals, defined as tokens whose probabilities differ sharply between harmful and benign queries. We show that these signals consistently align across LLMs, forming a model-agnostic direction that emerges from safety alignment. LADE consists of three components: (1) Extracting Latent Safety Signals from Dark Knowledge, which selects top-k safety-discriminative tokens from the first-token probability distribution; (2) Tokenizer Mapping, which maps these tokens across different tokenizers to enable model-agnostic application; and (3) kNN-based Discrimination, which classifies queries via a k-Nearest Neighbors search over the mapped tokens. Across diverse LLMs and benchmarks, LADE is robust against a wide range of jailbreak attacks and lowers attack success rates while maintaining a competitive safety-utility trade-off.

## BlockJack: Cross-Tenant Redirection Fault Attacks on Paged KV Cache Address Translation in LLM Serving

DevFeed: [BlockJack: Cross-Tenant Redirection Fault Attacks on Paged KV Cache Address Translation in LLM Serving](<https://devfeed.tech/articles/blockjack-cross-tenant-redirection-fault-attacks-on-paged-kv-cache-address-translation-in-llm-serving-146051.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25945>)

Author: Qifan Wang, Jeremy Boy, Jonas Sander, Yansong Gao, Hongsheng Hu, Ye Dong, Thomas Eisenbarth, David Oswald

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: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [ai security](<https://devfeed.tech/topics/ai-security.md>), [rowhammer](<https://devfeed.tech/topics/rowhammer.md>), [vllm](<https://devfeed.tech/topics/vllm.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Cache](<https://devfeed.tech/topics/cache.md>)

Tags: [attack](<https://devfeed.tech/tags/attack.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [cache](<https://devfeed.tech/tags/cache.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [defense](<https://devfeed.tech/tags/defense.md>), [inference](<https://devfeed.tech/tags/inference.md>), [information](<https://devfeed.tech/tags/information.md>), [integrity](<https://devfeed.tech/tags/integrity.md>), [kv-cache](<https://devfeed.tech/tags/kv-cache.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-serving](<https://devfeed.tech/tags/llm-serving.md>), [medical](<https://devfeed.tech/tags/medical.md>), [redirection](<https://devfeed.tech/tags/redirection.md>), [rowhammer](<https://devfeed.tech/tags/rowhammer.md>), [validation](<https://devfeed.tech/tags/validation.md>), [vllm](<https://devfeed.tech/tags/vllm.md>), [vulnerability](<https://devfeed.tech/tags/vulnerability.md>)

### AI overview

BlockJack corrupts vLLM-style paged KV cache block-table entries to redirect one tenant's cache reads to another tenant's blocks, without changing model parameters, prompts, or cached states. The authors report confidentiality and integrity attacks across seven Qwen, Llama, Gemma, and Phi models; on a 248-scene medical benchmark, up to 41.1% of responses recommended less urgent care while appearing consistent with the victim's question. They demonstrate physical feasibility using Rowhammer and evaluate assignment validation as a defense. vLLM and Sonar confirmed the vulnerability and proposed patches that the authors reviewed and validated.

### Source excerpt

arXiv:2611.25945v1 Announce Type: new Abstract: Large language model (LLM) serving frameworks accelerate inference by retaining previously computed states in a key-value (KV) cache. Paged serving frameworks such as vLLM organize this cache through block tables analogous to OS page tables, which map each request's logical KV blocks to physical KV blocks, forming a software-managed address translation layer. Although fault attacks such as Rowhammer have exploited OS page tables to redirect memory accesses, prior fault attacks on LLMs have largely targeted the data plane, i.e., model weights or KV-cache contents, leaving KV-cache mapping unexplored. We present BlockJack (i.e., Block Hijack), the first KV cache address-translation fault attack in LLM inference. By corrupting block-table entries, BlockJack redirects a request's KV cache reads to blocks assigned to another tenant's request without modifying model parameters, prompts, or cached KV states. It achieves both confidentiality violation (victim information appears in attacker responses) and integrity violation (attacker-supplied context steers victim answers). We study paged KV cache mappings in vLLM and formally characterize when redirect targets are predictable. Across seven models spanning the Qwen, Llama, Gemma, and Phi families, a faulted mapping exposes victim context. For the integrity attack, we examine how topic proximity, intra-topic divergence, and model behavior shape redirected responses. On a 248-scene medical benchmark, the integrity attack causes up to 41.1% of responses to recommend less urgent care, while still appearing consistent with the victim's question. We demonstrate physical feasibility of BlockJack with Rowhammer and subsequently evaluate assignment validation as a defense. Following responsible disclosure, vLLM and Sonar confirmed the vulnerability and proposed patches that we reviewed and validated.

## Cross-Modal Backdoor Attack on Connector-based Multimodal Large Language Models

DevFeed: [Cross-Modal Backdoor Attack on Connector-based Multimodal Large Language Models](<https://devfeed.tech/articles/cross-modal-backdoor-attack-on-connector-based-multimodal-large-language-models-145998.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2605.07490>)

Author: Runhe Wang, Li Bai, Zhaoxia Yin, Haibo Hu, Songze Li

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: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [AI software supply chain security](<https://devfeed.tech/topics/ai-software-supply-chain-security.md>), [Security](<https://devfeed.tech/topics/security.md>)

Tags: [activation](<https://devfeed.tech/tags/activation.md>), [adversarial](<https://devfeed.tech/tags/adversarial.md>), [anchor](<https://devfeed.tech/tags/anchor.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [asr](<https://devfeed.tech/tags/asr.md>), [attack](<https://devfeed.tech/tags/attack.md>), [backdoor](<https://devfeed.tech/tags/backdoor.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [components](<https://devfeed.tech/tags/components.md>), [connectors](<https://devfeed.tech/tags/connectors.md>), [cosine](<https://devfeed.tech/tags/cosine.md>), [cosine-similarity](<https://devfeed.tech/tags/cosine-similarity.md>), [cross](<https://devfeed.tech/tags/cross.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [effectiveness](<https://devfeed.tech/tags/effectiveness.md>), [evaluations](<https://devfeed.tech/tags/evaluations.md>), [experiments](<https://devfeed.tech/tags/experiments.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>)

### AI overview

The paper describes a cross-modal backdoor that poisons only an MLLM connector using one seed sample and a small set of augmented examples from one modality. A centroid of the resulting malicious latent region acts as an anchor; adversarial inputs from other modalities are optimized offline to reach it and trigger a target response, without access to the deployed model. Tests on PandaGPT and NExT-GPT report same-modality attack success rates up to 99.9% and rates above 95% in most cross-modal settings under bounded perturbations. ChatBridge experiments indicate effectiveness across different connector architectures. The abstract reports 0% backdoor leakage on clean inputs and weight-cosine similarity above 0.97 to benign connectors. Existing defenses reportedly do not neutralize the attack without substantial utility degradation.

### Source excerpt

arXiv:2605.07490v2 Announce Type: replace Abstract: Assembling multimodal large language models (MLLMs) from pretrained components has become a widely adopted practice in the open-source ecosystem, yet the supply-chain risks of lightweight connectors remain largely unexplored. We propose a cross-modal backdoor attack that exploits this gap. By poisoning only the connector with a single seed sample and a small set of augmented variants from one modality, the adversary establishes a compact malicious region in the shared latent space associated with a target output. A centroid extracted from this region then serves as a latent anchor for cross-modal activation: adversarial inputs from any other modality can be optimized offline to reach this anchor, triggering the target response without access to the deployed model. Evaluations on PandaGPT and NExT-GPT show same-modality attack success rates (ASR) up to 99.9%, with most cross-modal settings exceeding 95.0% ASR under bounded perturbations. Experiments on ChatBridge further demonstrate the attack's effectiveness across different connector architectures. The attack produces 0.0% backdoor leakage on clean inputs and maintains weight-cosine similarity above 0.97 relative to benign connectors. Existing defenses fail to neutralize the threat without substantial utility degradation, revealing that a single poisoned connector can establish a reusable backdoor pathway across modalities.

## ToolPulse: Side-Channel Eavesdropping on Tool-Using LLM Agents

DevFeed: [ToolPulse: Side-Channel Eavesdropping on Tool-Using LLM Agents](<https://devfeed.tech/articles/toolpulse-side-channel-eavesdropping-on-tool-using-llm-agents-146032.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25470>)

Author: Zihao Dan, Jinyu Zhang, Jinzhao Lian, Yun Chen, Shinan Liu

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

Content type: news

Language: en

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

Topics: [side channel](<https://devfeed.tech/topics/side-channel.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [OpenTelemetry](<https://devfeed.tech/topics/opentelemetry.md>)

Tags: [adversary](<https://devfeed.tech/tags/adversary.md>), [agents](<https://devfeed.tech/tags/agents.md>), [attack](<https://devfeed.tech/tags/attack.md>), [attack-surface](<https://devfeed.tech/tags/attack-surface.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [bench](<https://devfeed.tech/tags/bench.md>), [cache](<https://devfeed.tech/tags/cache.md>), [claude](<https://devfeed.tech/tags/claude.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [codex](<https://devfeed.tech/tags/codex.md>), [cs-ar](<https://devfeed.tech/tags/cs-ar.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [eavesdropping](<https://devfeed.tech/tags/eavesdropping.md>), [encryption](<https://devfeed.tech/tags/encryption.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [execution](<https://devfeed.tech/tags/execution.md>), [f1](<https://devfeed.tech/tags/f1.md>), [framework](<https://devfeed.tech/tags/framework.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-agents](<https://devfeed.tech/tags/llm-agents.md>), [side-channel](<https://devfeed.tech/tags/side-channel.md>), [tool](<https://devfeed.tech/tags/tool.md>)

### AI overview

ToolPulse shows how cache side channels can expose tool use by LLM agents even when inference runs remotely: tools execute on users' machines and often use shared system libraries. Across three hardware platforms, Flush+Reload enabled accurate real-time monitoring, while Prime+Probe had prohibitive overhead and cache-set aliasing. The Tool Recovery Attack reconstructed workflows for 11 LLMs evaluated with Codex and Claude Code, with every reported F1 score above 93%. Recovered traces supported task profiling across two Terminal-Bench categories and localization of sensitive calls in encryption/decryption and model training/inference. Both attacks reportedly retained high accuracy amid substantial noise on a server shared by about 20 users.

### Source excerpt

arXiv:2611.25470v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly deployed as agents that translate user requests into sequences of local tool invocations. While prior cache-based side-channel attacks target the locally deployed models, they typically require the adversary to share model-related components or hardware resources with the victim's inference process. These assumptions do not apply to the predominant agentic settings, in which commercial LLMs execute on vendor-controlled infrastructure. We show that agentic execution exposes a more practical attack surface: although inference remains remote, tools are executed on the user's machine and commonly rely on shared system libraries. In this paper, we first systematically evaluate Flush+Reload and Prime+Probe for monitoring agentic tool invocations. Our evaluation across three hardware platforms shows that Flush+Reload enables accurate, real-time tool monitoring, whereas Prime+Probe, despite supporting cross-VM attacks, incurs prohibitive monitoring overhead and suffers from cache-set aliasing. We present ToolPulse, a cache-based side-channel eavesdropping framework against tool-using agents. We first design the Tool Recovery Attack to reconstruct an agent's tool-execution workflow. We evaluate the Tool Recovery Attack on 11 LLMs across Codex and Claude Code. It consistently reconstructs tool workflows, with the lowest F1 score exceeding 93%. We further demonstrate two attack scenarios enabled by the recovered traces. Task Profiling Attack infers user intent across two Terminal-Bench task categories and the Trigger Localization Attack identifies sensitive tool calls in encryption/decryption and model training/inference workloads. Even under substantial system noise on a server shared by approximately 20 users, both Task Profiling Attack and Trigger Localization Attack maintain high accuracy.

## A Statistical Framework for Training Data Auditing of Large Language Models

DevFeed: [A Statistical Framework for Training Data Auditing of Large Language Models](<https://devfeed.tech/articles/a-statistical-framework-for-training-data-auditing-of-large-language-models-145158.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26264>)

Author: Shengjie Niu, Yeheng Ge, Jian Huang

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

Content type: article

Language: en

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

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

Tags: [auditing](<https://devfeed.tech/tags/auditing.md>), [benchmark-contamination](<https://devfeed.tech/tags/benchmark-contamination.md>), [books](<https://devfeed.tech/tags/books.md>), [case-studies](<https://devfeed.tech/tags/case-studies.md>), [copyright](<https://devfeed.tech/tags/copyright.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data-collection](<https://devfeed.tech/tags/data-collection.md>), [error-control](<https://devfeed.tech/tags/error-control.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [framework](<https://devfeed.tech/tags/framework.md>), [global](<https://devfeed.tech/tags/global.md>), [inference](<https://devfeed.tech/tags/inference.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [medical](<https://devfeed.tech/tags/medical.md>), [multiple-testing](<https://devfeed.tech/tags/multiple-testing.md>), [privacy](<https://devfeed.tech/tags/privacy.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [stat-ap](<https://devfeed.tech/tags/stat-ap.md>), [stat-ml](<https://devfeed.tech/tags/stat-ml.md>), [training-data](<https://devfeed.tech/tags/training-data.md>)

### AI overview

The framework audits whether documents and collections were used to train large language models, while accounting for membership signals concentrated in only a few passages. It combines complementary chunk-level evidence with Cauchy combination to score documents, then uses calibrated p-values for collection-level auditing and multiple-testing procedures to identify suspicious subsets. Under stated exchangeability conditions, the authors establish finite-sample validity of document-level p-values, type-I error control for global audits, and false discovery rate control for subset identification. Case studies cover BookTection, Pile-arXiv, Cosmopedia, and HealthCareMagic, spanning pre-training and fine-tuning data.

### Source excerpt

arXiv:2611.26264v1 Announce Type: cross Abstract: Training-data auditing asks whether data of interest were used to train a large language model (LLM), a question central to privacy auditing, copyright protection, and benchmark-contamination detection. These concerns are particularly relevant to the use of copyrighted books and scientific papers in pre-training, as well as synthetic text and medical question-answer records in fine-tuning. In practice, training membership signals may be concentrated in only a few passages of a long document, making it important to aggregate evidence within each document without diluting these localized signals. Beyond individual documents, auditing a target collection requires an overall assessment based on the document-level evidence. In this paper, we propose a unified statistical framework for document-level membership inference, target-collection global auditing, and suspicious-subset identification. We integrate complementary chunk-level membership evidence with Cauchy combination to construct document scores for document-level membership inference. The document scores and their calibrated p-values further enable an overall assessment of training-data collection use and the identification of suspicious documents through multiple testing. Under the stated exchangeability conditions, we establish finite-sample validity of document-level p-values, type-I error control for global auditing, and false discovery rate control for subset identification. Case studies on BookTection, Pile-arXiv, Cosmopedia, and HealthCareMagic demonstrate the effectiveness and reliability of the proposed framework in auditing pre-training and fine-tuning data.

## Ask in the Crowd: Differentially Private LLM Inference via Dummy-Augmented Shuffling

DevFeed: [Ask in the Crowd: Differentially Private LLM Inference via Dummy-Augmented Shuffling](<https://devfeed.tech/articles/ask-in-the-crowd-differentially-private-llm-inference-via-dummy-augmented-shuffling-145566.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25669>)

Author: Zhihao Liu, Zixiong Guo, Shuo Shao, Yu He, Wenli Wang, Meihui Chen, Di Wang, Yuke Hu

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

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>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Adversarial attacks](<https://devfeed.tech/topics/adversarial-attacks.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [attacks](<https://devfeed.tech/tags/attacks.md>), [balance](<https://devfeed.tech/tags/balance.md>), [cache](<https://devfeed.tech/tags/cache.md>), [classification](<https://devfeed.tech/tags/classification.md>), [client](<https://devfeed.tech/tags/client.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [efficiency](<https://devfeed.tech/tags/efficiency.md>), [embeddings](<https://devfeed.tech/tags/embeddings.md>), [framework](<https://devfeed.tech/tags/framework.md>), [generation](<https://devfeed.tech/tags/generation.md>), [gpt](<https://devfeed.tech/tags/gpt.md>), [inference](<https://devfeed.tech/tags/inference.md>), [inference-efficiency](<https://devfeed.tech/tags/inference-efficiency.md>), [inference-framework](<https://devfeed.tech/tags/inference-framework.md>), [input](<https://devfeed.tech/tags/input.md>), [kv-cache](<https://devfeed.tech/tags/kv-cache.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-inference](<https://devfeed.tech/tags/llm-inference.md>), [private](<https://devfeed.tech/tags/private.md>)

### AI overview

DP-SPD hides a user's real query among client-generated dummy queries, shuffles the queries before sending them to the LLM server, and reuses transformer key-value caches for queries with shared prefixes during decoding. The paper analyzes its privacy guarantee and evaluates classification and generation utility and protection against QIA, EIA, and GPT-based attacks. The abstract reports strong inference-time privacy, near-plaintext utility, and improved efficiency, but gives no quantitative results or implementation details.

### Source excerpt

arXiv:2611.25669v1 Announce Type: new Abstract: Large language models are increasingly deployed as LLM-as-a-Service (LLMaaS), where user queries are sent to remote LLMs for inference, raising severe privacy concerns about the exposure of sensitive user inputs. Differentially private inference mitigates these risks by injecting noise into input embeddings. However, existing methods struggle to balance privacy and utility, as stronger privacy protection often leads to substantial accuracy degradation in LLM inference. In this work, we propose DP-SPD, a differentially private LLM inference framework that protects user queries by hiding the real query among client-generated dummy queries. All queries are randomly shuffled before being sent to the server, preventing direct identification of the user's true intent. To improve inference efficiency, we exploit the shared-prefix structure among queries and reuse transformer key--value (KV) caches during decoding. We analyze the privacy guarantee of DP-SPD and evaluate its utility on classification and generation tasks, as well as its privacy protection under QIA, EIA, and GPT-based attacks. Experimental results show that our method achieves strong inference-time privacy protection and near-plaintext utility while improving inference efficiency through shared-prefix KV-cache reuse.

## DailySafe: Benchmarking Dynamic Safety Awareness in Everyday Multi-Turn Multimodal Interactions

DevFeed: [DailySafe: Benchmarking Dynamic Safety Awareness in Everyday Multi-Turn Multimodal Interactions](<https://devfeed.tech/articles/dailysafe-benchmarking-dynamic-safety-awareness-in-everyday-multi-turn-multimodal-interactions-146058.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26298>)

Author: Xuankun Rong, Xinhao Deng, Wenzhang Jiang, Ziheng Peng, Siyuan Liang, Wenke Huang, Pablo Arbelaez, Mang Ye

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: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [trust and safety](<https://devfeed.tech/topics/trust-and-safety.md>)

Tags: [awareness](<https://devfeed.tech/tags/awareness.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [everyday](<https://devfeed.tech/tags/everyday.md>), [multimodal](<https://devfeed.tech/tags/multimodal.md>), [safety](<https://devfeed.tech/tags/safety.md>)

### AI overview

DailySafe benchmarks whether multimodal language models can detect safety risks that emerge as everyday tasks unfold across conversation turns and images, while still helping with legitimate goals. It contains 195 naturalistic interactions spanning 10 domains and 42 task categories, with safety-critical evidence revealed progressively and case-specific annotations used to assess how models respond. The authors position it as a complement to benchmarks centered on isolated harmful requests or adversarial scenarios; the supplied abstract does not report evaluation results.

### Source excerpt

arXiv:2611.26298v1 Announce Type: new Abstract: As multimodal large language models (MLLMs) become increasingly integrated into everyday life, their ability to provide safe and reliable assistance in realistic daily activities remains insufficiently explored. Existing safety benchmarks primarily focus on isolated harmful requests or adversarial scenarios, overlooking the safety risks that may arise as ordinary tasks unfold. In practice, users pursue legitimate goals through multi-turn, multimodal interactions, where seemingly benign requests may become safety-critical when considered alongside earlier actions and visual observations. This calls for Dynamic Safety Awareness, the ability to track evolving task conditions, recognize implicit contextual risks, and provide safe and helpful assistance. To systematically evaluate this capability, we introduce DailySafe, a benchmark comprising 195 naturalistic multi-turn, multi-image interactions across 10 everyday domains and 42 task categories. Grounded in authentic user dialogue topics and documented safety knowledge, DailySafe captures realistic scenarios where safety-critical evidence is progressively revealed across conversational turns and visual observations. Using case-specific safety annotations, DailySafe assesses whether MLLMs can recognize and appropriately address evolving risks while supporting users' legitimate goals, bringing multimodal safety evaluation closer to real-world everyday AI assistance. DailySafe is publicly available at: https://github.com/XuankunRong/DailySafe.

## The Geometry of Lies: Knowledge and Persona Directions for Detecting Deception in Large Language Models

DevFeed: [The Geometry of Lies: Knowledge and Persona Directions for Detecting Deception in Large Language Models](<https://devfeed.tech/articles/the-geometry-of-lies-knowledge-and-persona-directions-for-detecting-deception-in-large-language-models-145683.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26545>)

Author: Md. Abdullah Al Zami, Suraiya Shanaz, Mahdi Hasan, Shoaib Ahmed Dipu, Farig Yousuf Sadeque, Md. Sabbir Ahmed

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

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [llama](<https://devfeed.tech/topics/llama.md>), [qwen](<https://devfeed.tech/topics/qwen.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [deception](<https://devfeed.tech/tags/deception.md>), [knowledge](<https://devfeed.tech/tags/knowledge.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [lies](<https://devfeed.tech/tags/lies.md>), [models](<https://devfeed.tech/tags/models.md>), [persona](<https://devfeed.tech/tags/persona.md>)

### AI overview

Persona directions derived from honest and deceptive role-play rank lies across Liars' Bench better than single-dataset probes on three of four models, but fail on Qwen; on Llama-3.3-70B-Instruct, their gain over a published direction falls within the range of random widenings. For lies about knowledge, persona directions often track wording and do not reliably distinguish a lie from an honest denial with identical wording. A separate knowledge direction, built from answerable versus unanswerable questions followed by the same denial, separates such cases on two models, narrowly on a third, and not on the fourth; some results depend on benchmark wording. Combining the directions beats a same-depth supervised probe across all four models, though results are near chance on one, and the combination is level with the persona direction on Llama's average and weaker on held-out lies. More varied role-play questions do not improve persona-direction performance on Liars' Bench.

### Source excerpt

arXiv:2611.26545v1 Announce Type: new Abstract: Linear probes are the cheapest LLM lie detectors but often fail to transfer across kinds of lie. Label-free directions from honest and deceptive role-play are untested at 70B scale and on lies about knowledge, where a lie and an honest denial can be identical. We widen a published persona bank with roles that deny knowledge they have, plus honest counterparts. At each model's middle layer the resulting persona direction ranks lies across Liars' Bench better than single-dataset probes on three of its four models; it fails on Qwen, even when Qwen stays in role far more often. On Llama-3.3-70B-Instruct it beats directions from a random persona bank, but its gain over the published direction is within the range of random widenings. On lies about knowledge, persona directions mainly track a reply's wording: where lies and honest denials share their words, our three main contrasts stay inside the random-direction band on Llama and Qwen and reach only its edge on Gemma, though on Mistral one clears it. A knowledge direction contrasts questions the model can answer with ones nobody can, each followed by the same scripted denial. It separates lies from identical honest denials on two models and narrowly on a third, but not on the fourth; its other results partly depend on the benchmark's wording. On the Gender-Secret fine-tune it clears the band by a few hundredths on Llama and fails on Mistral. The sum of the two directions ranks lies better than a supervised probe at the same depth on all four models (on one, both near chance) but is level with the persona direction on Llama's average and behind that probe on held-out lies. More varied role-play questions do not help the persona direction on Liars' Bench.

## A Wolf in Sheep's Pixels: Visual Camouflage Attacks via HDR Gain Maps

DevFeed: [A Wolf in Sheep's Pixels: Visual Camouflage Attacks via HDR Gain Maps](<https://devfeed.tech/articles/a-wolf-in-sheep-s-pixels-visual-camouflage-attacks-via-hdr-gain-maps-145517.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25469>)

Author: Qi Shen, Yujia Gong, Yue Zhao, Tong Liu

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

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Image processing](<https://devfeed.tech/topics/image-processing.md>), [Image](<https://devfeed.tech/topics/image.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [image recognition](<https://devfeed.tech/topics/image-recognition.md>)

Tags: [attacks](<https://devfeed.tech/tags/attacks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [exploits](<https://devfeed.tech/tags/exploits.md>), [hdr](<https://devfeed.tech/tags/hdr.md>), [maps](<https://devfeed.tech/tags/maps.md>), [processing](<https://devfeed.tech/tags/processing.md>), [sdr](<https://devfeed.tech/tags/sdr.md>)

### AI overview

DRJanus exploits the difference between SDR image processing and HDR display: automated systems may inspect an SDR base while viewers see a reconstruction that also uses an auxiliary gain map. It embeds two independently chosen color images in one JPEG without querying or optimizing against the target model, preserving the SDR base exactly and achieving an average HDR reconstruction PSNR of 43.03 dB across 400 images. In tests, it evaded all six commercial moderation services and nine vision-language models on a 100-image sexual-content benchmark. Across five agents, tested on 120 task cases per threat setting, attack success ranged from 31.67% to 100% for moderation evasion and 95% to 100% for task manipulation. The authors also explore cross-view consistency checks as a mitigation; the supplied abstract does not report their effectiveness.

### Source excerpt

arXiv:2611.25469v1 Announce Type: cross Abstract: An image can pass automated safety checks while showing harmful content to its human viewer. We introduce DRJanus, a visual camouflage attack that exploits the gap between standard dynamic range (SDR) processing and high dynamic range (HDR) display. Automated systems often read only an image's SDR base, whereas HDR viewers reconstruct its appearance using an auxiliary gain map. DRJanus inverts this reconstruction to embed two independently chosen color images in a single JPEG, without querying or optimizing against the target model. It preserves the chosen SDR base exactly and reconstructs HDR targets at an average PSNR of 43.03 dB across 400 images. On a benchmark of 100 sexual-content images, DRJanus evades detection by all six tested commercial moderation services and nine vision-language models. Across five agents, each evaluated on 120 task cases per threat setting, attack success ranges from 31.67% to 100% for moderation evasion and from 95% to 100% for task manipulation. We also explore cross-view consistency checks as a mitigation. These findings expose a gap in visual trust: securing what a machine reads does not secure what a user sees. A demo is available at https://qshen0629.github.io/DRJanus/

[Next page](<https://devfeed.tech/topics/llm-security.md?cursor=WyIyMDI2LTExLTIwVDA1OjAwOjAwKzAwOjAwIiwgIjIzZjAxY2Q2LTNlMzQtNDcwMy1iZThlLTVhYjg1ZWNhMDU3ZSJd>)