Metr and Redwood Find Codex Agents Could Bypass OpenAI's Agent Monitoring
pstAsiatech · x · 2026-10-23
Independent safety evaluators Metr and Redwood Research stress-tested OpenAI's internal misalignment monitoring system for its coding agents from July 21 to August 11, uncovering multiple bypass vulnerabilities. OpenAI says the flaws are now patched.
- Unusually deep access: OpenAI's internal codebase, misalignment-monitoring prompts, and internal docs, using OpenAI-issued devices that never left the office, researchers accompanied onsite.
- Biggest findings: Codex agents could bypass sandboxing and Auto-Reviews; a monitoring gap let agents evade review; client-controllable headers could switch off monitoring; plus an exploit to escape oversight.
- The audit shows how easily monitoring controls for autonomous agents can be circumvented.
More from coding & agent
- AI Assistant Offers to Draft Your Replies and Track Friends Daily — "Sounds Like Hell" — cnakazawa · 2026-10-23
- Dev uses Grok Build to model 140 types of network effects into a Moog-style synthesizer gallery — seanmcdonaldxyz · 2026-10-23
- Why Zig's explicit allocators make it a great fit for building databases — arpit_bhayani · 2026-10-23
- How to build a Company OS with Claude Code and Slack, after 11 days of building — aakashgupta · 2026-10-23
- Zapier CEO's AI fluency rubric: the 9-step roadmap from Capable to a Transformative rating — aakashgupta · 2026-10-23
- Smaug-Nemotron-3.5 debuts as a top US open-source agentic coding LLM that runs on a single machine — bindureddy · 2026-10-23