# How Agent Anomaly Detection scores on benchmarks we didn't write

DevFeed: [How Agent Anomaly Detection scores on benchmarks we didn't write](<https://devfeed.tech/articles/how-agent-anomaly-detection-scores-on-benchmarks-we-didn-t-write-131280.md>)

Original publisher: [Read original article](<https://developers.googleblog.com/how-agent-anomaly-detection-scores-on-benchmarks-we-didnt-write/>)

Author: Nakul Javali

Published: 2026-11-13T06:38:11.586756Z

Content type: news

Language: en

Sources: [Google Developers Blog](<https://devfeed.tech/sources/google-developers-blog.md>)

Topics: [threat detection](<https://devfeed.tech/topics/threat-detection.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [agent observability](<https://devfeed.tech/topics/agent-observability.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [google-adk](<https://devfeed.tech/topics/google-adk.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agentic](<https://devfeed.tech/tags/agentic.md>), [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [anomaly-detection](<https://devfeed.tech/tags/anomaly-detection.md>), [asynchronous](<https://devfeed.tech/tags/asynchronous.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [defense](<https://devfeed.tech/tags/defense.md>), [defense-in-depth](<https://devfeed.tech/tags/defense-in-depth.md>), [detection](<https://devfeed.tech/tags/detection.md>), [instructions](<https://devfeed.tech/tags/instructions.md>), [latency](<https://devfeed.tech/tags/latency.md>), [logs](<https://devfeed.tech/tags/logs.md>), [public](<https://devfeed.tech/tags/public.md>), [security](<https://devfeed.tech/tags/security.md>), [tool](<https://devfeed.tech/tags/tool.md>), [traces](<https://devfeed.tech/tags/traces.md>)

## AI overview

In five public benchmarks totaling 1,774 sessions, Agent Anomaly Detection (AAD) flagged 87% of harmful instructions that agents carried out, analyzing traces and logs asynchronously without adding runtime latency. Detection varied by benchmark (85% on AgentDojo, 87% on InjecAgent, 99% on AgentDyn, and 81% on AgentHarm); across all harmful sessions, including those agents never read or resisted, it flagged 45%-62%. Benign-session false alarms ranged from 2%-4% on office-task injection benchmarks to 16% on AgentHarm, whose benign tasks used high-impact tools. AAD missed most resisted instructions and caught only 4 of 18 cases where poisoned long-term memory led to harmful action. Results used one ADK agent model and simulated apps; production false-alarm rates may differ, and AAD returned no result for about 6% of sessions. The source recommends combining detection with controls that prevent harmful actions.

## Source excerpt

Agent Anomaly Detection (AAD) provides asynchronous, low-latency monitoring for AI agents by analyzing OpenTelemetry traces and logs to flag goal hijacks and tool misuse. Tested across five public security benchmarks, AAD successfully caught 87% of harmful instructions carried out by agents without adding runtime latency. Developers can integrate AAD alongside semantic governance and strict tool limits as part of a defense-in-depth strategy to secure agentic applications.