Google Wants to Catch Agents That “Succeed” the Wrong Way
Google’s new anomaly detector audits how enterprise agents reason and use tools. Task completion alone is no longer enough.
Google’s Agent Anomaly Detection adds a behavioral audit layer for enterprise agents. Now in private preview on Gemini Enterprise Agent Platform, it reads OpenTelemetry logs and traces across a session, then checks reasoning paths, tool calls and execution flow for suspicious intent, policy violations and unusual behavior.
The service combines lightweight outlier detection with deeper LLM analysis and tool-invocation reconstruction. Findings include severity, an explanation and recommended action, and surface in Security Command Center. An API lets teams set thresholds that can stop later tool calls or halt the next turn.
For PMs, this widens the definition of “works.” An agent can complete a task while using the wrong tool, exceeding its authority or consuming unreasonable resources. Launch criteria need behavioral boundaries, trace coverage, a clear incident owner and a response path alongside quality, latency and cost.
The limits matter. This is an opt-in private preview for Google Agent Runtime deployments using Python ADK, with US multi-region telemetry requirements. Analysis is asynchronous, so it may not stop the first harmful action. Google has not published accuracy or false-positive benchmarks, and custom business detectors are planned rather than available.
