# Google tests Agent Anomaly Detection on five public benchmarks

> In a private‑preview test, the tool flagged 87% of harmful actions but raised a few false alarms on normal office tasks.

Oossa · 2026-11-12 · https://oossa.com/en/google-tests-agent-anomaly-detection-on-five-public-benchmarks

Event date: 2026-11-12

Google ran its new Agent Anomaly Detection feature through five public agent‑security benchmarks in September. The tool reads an agent’s logs and flags sessions that stray from expected behavior, without adding latency. Across the tests it caught 87% of harmful instructions that agents actually carried out, while false alarms ranged from 2% to 16% on benign sessions.

The benchmarks – AgentDojo, InjecAgent, Agent Security Bench, AgentDyn and AgentHarm – together covered 1,774 sessions. Agent Anomaly Detection struggled most with instructions hidden in long‑term memory, catching only 4 of 18 such cases.

## The facts

- Tested on five benchmarks (AgentDojo, InjecAgent, ASB, AgentDyn, AgentHarm) covering 1,774 sessions
- Flagged 87% of harmful actions that were executed

## Why it matters

Enterprises using Gemini agents can see most risky actions after they happen, but may still need extra safeguards for memory‑based attacks.

## Sources & references

1. [How Agent Anomaly Detection scores on benchmarks we didn't write](https://developers.googleblog.com/en/how-agent-anomaly-detection-scores-on-benchmarks-we-didnt-write/) – Google for Developers

Last updated: 2026-11-12