Three-week on-site audit finds OpenAI leaves 25% of frontier agent traffic unmonitored
The Decoder · rss · 2026-10-24
External auditors from METR and Redwood Research spent three weeks on-site at OpenAI examining its internal monitoring system for AI agents — and found major gaps: up to 25% of agent traffic from models at the GPT-5.5 class and above goes unmonitored. OpenAI heavily redacted the 61-page report; only a summary is public. The audit is part of OpenAI's external oversight under its frontier safety framework, but the redactions and the gap numbers expose real weaknesses in its internal safeguards.
More from Safety
- For brands building AI assistants: make customer memories visible and deletable — petesena · 2026-10-24
- How to audit your ChatGPT memory tonight — plus 'Attribution Shield' fix — petesena · 2026-10-24
- User asks ChatGPT to make their job 'more interesting' — it files a fake career as memory — petesena · 2026-10-24
- Real ChatGPT memories read like therapy notes: loneliness, fear of failure — petesena · 2026-10-24
- OpenAI to showcase the next era of cyber defense on November 12 — imjustnewatai · 2026-10-24
- Monthly CVEs explode toward 17,000, roughly 10x the 2018 baseline — burny_tech · 2026-10-24