StudioMETR and Redwood researchers red-team OpenAI's internal agent monitoring and find multiple bypasses
Original title: Two outsiders found ways OpenAI’s AI agents could slip past the system built to catch them
METR reported on October 22 that a METR employee and a Redwood Research employee spent three weeks inside OpenAI from July 21 to August 11 and found several novel vulnerabilities in its internal misalignment monitoring of coding agents like Codex, including two headers that start an unmonitored agent copy, an exploit that kept hundreds of agent actions out of the monitor's sight, and a flaw letting agents indefinitely postpone monitoring, which left roughly 10% of eligible Codex traffic unmonitored from May to July. METR estimated 75 to 94% of agentic traffic on the internal Responses API from models at least as strong as GPT-5.5 was eligible for monitoring, and 99.2% for internal Codex agents alone. OpenAI says every vulnerability affecting its production safeguards has been patched, but…
The article lays out the specific vulnerabilities found and OpenAI's remediation claims, giving readers a concrete look at the real boundaries of agent monitoring coverage.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.