Understand this story
On November 12, 2026, Google published the test results of its Agent Anomaly Detection on five public agent security benchmarks not written by Google. The evaluation covered 1,774 sessions, with a flagging rate of 87% for harmful instructions actually executed by agents and a false positive rate of 2% to 16% for benign sessions. Broken down by benchmark, the flagging rate for executing harmful instructions was 85% on AgentDojo, 87% on InjecAgent, 99% on AgentDyn, and 81% on AgentHarm. Google stated that the feature is strongest against goal hijacking and tool misuse, and weakest against memory poisoning, catching only 4 out of 18 cases. There are currently no further developments.
AI-generated from public reports
Latest development · Report
How does Agent Anomaly Detection perform on five benchmarks not written by Google?Developments
1 confirmed developments
Google publishes test results of Agent Anomaly Detection on five public benchmarks
Source:Google Developers Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 12
Google Developers BlogFirst-party
How does Agent Anomaly Detection perform on five benchmarks not written by Google?
Google's Agent Anomaly Detection evaluated 1,774 sessions across five publicly available agent security benchmarks not written by Google, flagging 87% of harmful instructions that agents actually executed, with false positive rates of 2% to 16% on benign sessions. Broken down by benchmark, the flagging rate for executed harmful instructions was 85% on AgentDojo, 87% on InjecAgent, 99% on AgentDyn, and 81% on AgentHarm. The feature is strongest against goal hijacking and tool misuse, and weakest against poisoned memory, catching only 4 out of 18 cases.