Google ran its new Agent Anomaly Detection feature through five public agent‑security benchmarks in September. The tool reads an agent’s logs and flags sessions that stray from expected behavior, without adding latency. Across the tests it caught 87% of harmful instructions that agents actually carried out, while false alarms ranged from 2% to 16% on benign sessions.
The benchmarks – AgentDojo, InjecAgent, Agent Security Bench, AgentDyn and AgentHarm – together covered 1,774 sessions. Agent Anomaly Detection struggled most with instructions hidden in long‑term memory, catching only 4 of 18 such cases.
Why it matters
Enterprises using Gemini agents can see most risky actions after they happen, but may still need extra safeguards for memory‑based attacks.
