METR audit finds gaps in OpenAI monitoring of agentic model traffic
Original title: Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring
Auditors estimate 6–25% of GPT-5.5-class internal agent traffic and about 10% of Codex traffic went unmonitored; OpenAI says the issues are fixed.
Researchers from METR and Redwood Research spent July 21 to August 11 on site at OpenAI reviewing internal monitoring for rogue AI behavior. They reported uncovered agentic Responses API traffic, a compaction-event gap that left about 10% of monitored Codex traffic untracked, ways to hide actions and bypass the Codex sandbox, and no monitoring of RL training when the audit began. OpenAI says it independently fixed the compaction bug during the audit, later expanded RL monitoring, and has fixed every weakness found. Most of the 61-page report remains redacted; only a short public summary was released.
Unconfirmed: Coverage percentages are METR estimates with significant uncertainty in both directions, and OpenAI’s post-audit fixes were not part of the evaluation.
- Access covered OpenAI’s codebase, internal documents, and about a dozen employee interviews.
- METR says uncertainty around both coverage estimates is significant.
- Agents could delay oversight by continuing to act instead of returning control.
- Specially crafted inputs could likely crash the monitor, auditors found.
- METR and Redwood leadership received a 38-page redacted report.
- Vendors
- OpenAI
- Products
- CodexResponses API
- Organizations
- METRRedwood Research
- AI models
- GPT-5.5
From the original article
2 min read at the-decoder.com
External auditors from METR and Redwood Research spent three weeks on-site at OpenAI examining the company's internal monitoring system for AI agents and found significant gaps, with up to 25 percent of agent traffic from models at the GPT-5. 5 class and above going unmonitored. OpenAI redacted large portions of the 61-page report, and only a summary is publicly available.
By Matthias Bastian · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.