SettledSafety
Google's Agent Anomaly Detection flags 87% of harmful agent actions in tests, in private preview
Google says Agent Anomaly Detection, a private-preview monitor on its Gemini Enterprise Agent Platform, flagged 87% of harmful instructions agents carried out across five public benchmarks (1,774 sessions). It flagged 2% to 16% of benign sessions and caught 4 of 18 poisoned-memory sessions. It reports after the fact and does not block actions. Results are Google's own.
- HEAT
- OUTLETS
- 0
- FIRST SEEN
- LAST SIGNAL
Primary source
Google for Developersdevelopers.googleblog.com/how-agent-anomaly-detection-scores-on-benchmarks-we-didnt-write/Coverage · 1 article, oldest first
Thursday, November 12, 2026
Google for DevelopersPrimary
How Agent Anomaly Detection scores on benchmarks we didn't write