Agent Anomaly Detection scores on benchmarks we didn't write
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Google tested its Agent Anomaly Detection feature on five public, non-Google agent-security benchmarks covering prompt injection and direct harmful requests. The feature flagged 87% of harmful instructions that agents actually carried out, with 2% to 16% false alarms on benign sessions.