METR-Redwood red team found ~10% of OpenAI's internal Codex traffic went unmonitored
dl_weekly · x · 2026-11-04
Redwood Research published a write-up of a rare high-access red-team exercise: from July 21 to Aug 11, 2026, one METR and one Redwood employee spent three weeks stress-testing OpenAI's internal misalignment monitoring system, producing a 61-page internal report.
Key facts:
- Compaction events left roughly 10% of eligible internal Codex traffic unmonitored
- The report also assessed whether agents could in principle disable or subvert the monitoring system
- Unusually deep access: OpenAI's internal codebase, monitoring prompts/config, internal docs, plus interviews with a dozen researchers across monitoring, safety-systems, and security teams; work happened on OpenAI-issued devices that never left site
- All three parties framed it as a pilot for external deep access, not public transparency — hence only a brief high-level public post; future versions promise stronger independence and transparency practices
Also posted on METR's blog; author Buck Shlegeris (Redwood Research).
More from Companies & People
- ThursdAI: Gemini 4 Carbon takes #1 with 59 on AA index, Anthropic files S-1 IPO — thursdai_pod · 2026-11-04
- fal Co-founders on Generative Media Rise and AI Infra Economics — gorkem · 2026-11-04
- Unusual wave of public departures hits Anthropic right before IPO, founder notes — zeeg · 2026-11-04
- Canva restructuring hits 1,271 staff as nearly a quarter face layoffs or redeployment — CtrlAltDwayne · 2026-11-04
- Sam Altman on OpenAI exec exodus: 'They only trust themselves to build this' — ChrisGPT · 2026-11-04
- Vorelios (YC F26) launches AI that predicts physics in seconds to engineer designs autonomously — KlausCodes · 2026-11-04