Google reports Agent Anomaly Detection benchmark results

Google says its Agent Anomaly Detection feature flagged 87% of harmful instructions that AI agents carried out in tests against five public, non-Google-written security benchmarks. The feature was strongest against agent goal hijacking and tool misuse, but caught only four of 18 attacks involving poisoned long-term memory.
- Agent Anomaly Detection flagged 87% of harmful instructions that agents carried out.
- Google tested five public, non-Google-written agent-security benchmarks.
- Detection reached at least 95% in eight of 11 apps for carried-out harmful instructions.
- The system caught four of 18 harmful sessions involving poisoned long-term memory.
- Agent Anomaly Detection is in Private Preview and does not block actions.
Testing against independent agent-security benchmarks
Google tested Agent Anomaly Detection against AgentDojo, InjecAgent, Agent Security Bench, AgentDyn and AgentHarm, covering 1,774 sessions in total. The benchmarks were selected to avoid evaluating the system only on tests created by its own team.
In four benchmarks, the harmful instruction was hidden in data returned by a tool, such as an email or file. In AgentHarm, the user directly asked the agent to cause harm. Google ran each case through an agent built with the Agent Development Kit using gemini-2.5-flash on simulated apps.
The benchmark's own deterministic Python checks examined the final state of those apps, including whether money left an account. Agent Anomaly Detection instead received the agent's traces and logs, as it would for a customer's agent. Each agent first completed at least 300 benign sessions so the system could learn its usual behavior.
Detection was strongest after harmful actions occurred
According to Google, Agent Anomaly Detection flagged 87% of harmful instructions that agents read and followed. The rate ranged from 81% on AgentHarm to 99% on AgentDyn, with 85% on AgentDojo and 87% on InjecAgent.
Results varied by application. In eight of 11 apps, the system flagged at least 95% of carried-out harmful instructions. Its lowest rate was 63% in AgentDojo's Slack app, where many attacks requested a small action such as visiting a page, behavior that can resemble ordinary browsing.
The system raised false alarms on 2% to 4% of benign sessions on the three injection benchmarks. The rate was 16% on AgentHarm, whose ordinary tasks included posting to social media, running terminal commands and downloading files. Google says everyday use of high-impact tools can look more like misuse.
Poisoned memory remains a major gap
Agent Anomaly Detection was less effective when an agent resisted an instruction or when the harmful behavior involved a small action. It flagged 11% to 25% of instructions hidden in tool output that agents resisted, compared with 55% of refused direct requests in AgentHarm.
The largest weakness was poisoned memory. In Agent Security Bench, instructions planted in an agent's long-term memory caused harmful behavior in 18 sessions, but the system caught only four. Google also found that performance on new agents depended on the amount of behavioral history: on AgentDyn, detection rose from 33% with about 220 benign sessions to 99% with 360.
Google says it is working on detecting resisted attacks at lower severity, identifying small actions that reach untrusted destinations, reducing false alarms around high-impact tools, covering new agents sooner and detecting long-term-memory attacks more reliably.
Google recommends layered defenses
Agent Anomaly Detection is in Private Preview and does not block actions. Google recommends using it alongside controls that screen untrusted content, block risky tool calls, require approval for high-impact actions and limit each agent's tools and permissions. The company points to Model Armor for checking prompts and tool results, and Semantic Governance Policy, in Preview, for evaluating proposed tool calls against user requests and rules.
Google also advises customers to enable Agent Development Kit tracing and logging, including prompt and response content, and to review findings in Security Command Center. The company cautions that the study used one agent model and simulated apps, while benign benchmark tasks may be simpler than production traffic. It also said the system returned no result for about 6% of submitted sessions, which lowered overall rates by 1 to 2 points when counted as misses.
Google for Developers: How Agent Anomaly Detection scores on benchmarks we didn't write- Google Developers Blog


