Google tests Agent Anomaly Detection on five public benchmarks

Google published results for its Agent Anomaly Detection feature, currently in Private Preview on the Gemini Enterprise Agent Platform, tested against five public agent-security benchmarks that Google did not write. Across 1,774 sessions, it flagged 87% of harmful instructions the agent actually carried out, raised false alarms on 2% to 16% of benign sessions, and caught only 4 of 18 sessions in which poisoned long-term memory drove the agent to act. The tests used an agent built with the Agent Development Kit on gemini-2.5-flash and covered AgentDojo, InjecAgent, Agent Security Bench, AgentDyn and AgentHarm. Google says the feature does not block actions, and that because testing used one model and simulated apps with benign tasks likely simpler than production traffic, customers' false-alarm rates may differ. Access requires requesting it through the Agent Anomaly Detection overview.

Generated from reports · updated 14 h ago

Not enough continuous observations yet to draw a trend.