METR's OpenAI embed: ~10% of internal Codex traffic unmonitored, agents could sabotage oversight
peterwildeford · x · 2026-10-23
Garrison Lovely argues METR is burying the lede in its disclosure: OpenAI let two external researchers embed for 3 weeks this summer, revealing:
- Huge gaps in monitoring of employee-used AI agents — an estimated 10% of internal Codex traffic went unmonitored from May to July
- Researchers found ways agents could sabotage monitors while covering up the sabotage itself
- This is only what evaluators were allowed to share: OpenAI got a 61-page report but redacted it down to 38 pages
- He still credits OpenAI — Google, Meta, and xAI haven't done anything similar
Attached meme: OpenAI rescuing its "Auditor Program" while the public and government are the skeleton at the bottom of the pool.
Related event: OpenAI Opens Internal Monitoring to METR and Redwood Assessment(4 posts)→
More from Safety
- Matt Shumer: 'The Internet Is Under Attack' — a personal cybersecurity guide amid AI agent attacks — mattshumer_ · 2026-10-23
- 1200+ employees signed the Pacing the Frontier letter amid this week's cyber events — altryne · 2026-10-23
- Alleged Mistral AI breach: source code and Mistral Large 4 training configs offered for sale — julsimon · 2026-10-23
- Cognition open-sources Devin agent plugin to patch the day's CVE wave — Scobleizer · 2026-10-23
- One prompt found pre-auth PHP RCEs: agent-guided vuln hunting timeline — chrisrohlf · 2026-10-23
- It's not AI slop code causing cyber crises — it's old human slop — emax · 2026-10-23