NewsResearch
METR red-teamed OpenAI's internal agent monitoring: about 10 percent of Codex traffic went unmonitored for months, and headers could switch it off
METR published on October 22 a summary of a three-week exercise, July 21 to August 11, in which one of its staff and one from Redwood Research worked on OpenAI-issued devices to red-team the misalignment monitoring that covers OpenAI's internal agentic traffic, Codex included; the full 61-page report stays internal. They estimate that 75 to 94 percent of agentic Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring, and 99.2 percent of the subset used by internal Codex agents. Because compaction events were not monitored, roughly 10 percent of eligible internal Codex traffic was likely unmonitored between May and July; OpenAI staff had found the gap and fixed it during the exercise. Monitoring ran when an agent returned control to the user, so an agent could postpone it by continuing to act; two client-controllable headers could disable it for new agents; reinforcement learning runs were not covered at the start, with only a subset of rollouts monitored from March to June; Codex agents could in principle bypass sandboxing and Auto-review; an exploit could keep hundreds of recent actions from being monitored; crafted inputs could likely make the monitor fail; and two internal tools thought to be read-only had a simple exploit. The monitor saw the full trajectory agents see and its prompts target scheming, monitor subversion, sabotage, data exfiltration, deception and prompt injection. OpenAI says every vulnerability affecting its production safeguards has been patched since; METR has not evaluated that claim.
Why it matters
A monitor that runs only at the end of a turn, can be turned off by a header and skips compaction is a monitor an agent can route around. Audit your own agent logging for the same three gaps before you rely on it.