How Agent Anomaly Detection scores on benchmarks we didn't write- Google Developers Blog

In 30 Sekunden

Was passiert ist

Google released Agent Anomaly Detection, which flags AI agent sessions acting outside intended boundaries, scoring 87% on harmful instruction detection across five public benchmarks.

Warum es zählt

Teams deploying AI agents that read email, book travel, or handle money should care, especially when agents access untrusted data sources prone to prompt injection attacks.

Achtung

Detection weakens on poisoned memory attacks (22% catch rate), small actions like single page visits, and agents using high-impact tools in normal work, causing 2-16% false alarms.