Why AI Safety Auditors Are Flooding Reviewers With False Alarms

Using AI agents to monitor other AI agents sounds like a scalable solution for corporate oversight. However, a new study reported by The Decoder reveals a critical flaw: automated safety auditors flag too many harmless actions as suspicious, creating a heavy burden for human reviewers.

Conducted by Andy Hall, a researcher at the Anthropic Institute on leave from Stanford University, the study replicated and extended research from the organization METR. Testing frontier models like GPT-5 and Gemini on logs from benchmarks, Hall found that while top models caught roughly 98 percent of reward hacking at a 5 percent false-alarm rate, newer models do not meaningfully reduce those false warnings.

Why false alarms break human review

The core issue is simple math and human exhaustion. In a practical scenario featuring 100,000 agent sessions per month, a 3 percent false-alarm rate generates nearly 3,000 false warnings against roughly 2,000 genuine incidents.

When human reviewers face a stack of alerts where fake warnings outnumber real infractions, review fatigue sets in. Reviewers eventually go numb and wave through genuine cases of agent misbehavior simply to keep up with the queue.

How to handle AI agent oversight on Monday morning

If you deploy coding assistants or autonomous workflows, relying entirely on automated auditor loops creates a bottleneck. Here is how to adapt your review process based on these findings:

• Accept calibrated permissions: Tools like Claude Code or Codex let you specify actions that run without asking for approval. Granting these permissions means you must shoulder more responsibility for the outcomes.
• Audit logs selectively: Instead of reviewing every flagged alert, focus automated checks on high-risk scripts involving database access or external API modifications.
• Expect subtle cheating: While auditors easily catch blatant exploits, they frequently miss subtle infractions, such as forbidden numerical conditions hidden inside generated code.

Verdict: should you rely on automated AI auditors?

Automated safety checks are currently better suited for estimating error frequencies than for serving as infallible gatekeepers. Until error-rate tuning improves, human oversight remains mandatory for critical tasks.