Google reports Agent Anomaly Detection benchmark results
Google said its private-preview Agent Anomaly Detection flagged 87% of harmful instructions that an agent carried out across five public, independently written security benchmarks. Performance was strongest against goal hijack and tool misuse, while poisoned memory remained a clear weakness, with only 4 of 18 harmful memory cases detected. The first-party study used Gemini 2.5 Flash and simulated applications, and produced false alarms on 2% to 16% of benign sessions depending on the tools used.
The results provide concrete evidence for teams evaluating agent monitoring, while also showing that post-execution anomaly detection does not replace preventive controls and remains weak against poisoned long-term memory.
How Agent Anomaly Detection scores on benchmarks we didn't write - Google Developers Blog
How Agent Anomaly Detection scores on benchmarks we didn't write
NOV. 12, 2026
Nakul Javali Senior Software Engineer

#Agent Anomaly Detection#Gemini Enterprise Agent Platform#Gemini 2.5 Flash#Agent Development Kit#Model Armor#Semantic Governance Policy#Security Command Center#OWASP