METR red team broke OpenAI's agent monitoring in three ways

Oct 26, 2026 · Issue 13 · 1 source, 1 article

METR's red team found that the Codex sandbox could be bypassed, that hundreds of actions could be hidden from the monitor, and that serialization mismatches could disable the monitor itself. The results attack the basic assumptions of agent oversight.

Who is affected and how

CTO

Sandboxing and logging protections for agents require complementary controls.

Narrative: Control slips away: deleting and autonomous agents

Sources

• METR’s Red Team Broke OpenAI’s Agent Monitoring Three Ways. Each One Attacks a Different Assumption.forkast.news · Oct 26, 2026

Machine-generated summary based on the titles and excerpts of the source articles. The facts come from the linked article.