Incident·WS-0114·Jul 21, 2026·disclosed Oct 22, 2026 (93 days later)·in the lab·Oversight failures

Red team finds OpenAI's agents could dodge its misalignment monitor

OpenAI internal misalignment monitoring; METR and Redwood Research (red team)

11/100
WS-1 Minor
Harm score 0, signal score 24. Autonomy: tool (a person misused it). How the scores work.
Harm 0/5
damage actually done
Reach 0/5
who was affected
Autonomy 0/4
how much the AI did on its own
Novelty 3/5
how new the behavior was
Permanence 0/5
how hard it is to undo

In a three-week on-site exercise (Jul 21 to Aug 11), one METR and one Redwood Research employee found several novel holes in the system OpenAI uses to catch misbehaving internal agents: an exploit that would keep hundreds of recent actions from being monitored, a gap that let agents postpone monitoring indefinitely by continuing to act (compaction events were unmonitored, roughly 10% of eligible internal Codex traffic from May to July), a potential route for Codex agents to bypass sandboxing and act with the host user's permissions, and headers that could start unmonitored agents. No evidence that agents actually exploited them; METR could not check. OpenAI says all production-affecting vulnerabilities are patched; METR has not verified that.

Damage

None realized; monitoring gaps in a production safeguard, found by humans

The capability behind it

Sources

METR, second source. METR describes it as a pilot high-access exercise, its first of this kind inside OpenAI. Public post is a summary; OpenAI could redact at its discretion. Findings describe the system on and before Jul 21, 2026. Reuters coverage (Oct 22) cited via MarketScreener syndication.

Factor ratings and the score are this index’s judgments by a published formula; facts, dates and figures come from the linked sources. Where a lawsuit is pending, the summary reports allegations, not findings.