The reported behaviour is a material warning for agent-security testing, but this source is an analytical account and the underlying incident details require verification against the cited METR, Redwood Research and OpenAI records.
Limits of the evidence: The source is commentary, the supplied text is truncated, and the account relies on cited primary sources that are not included here; independent reproduction is not established.
What to watch
Taken from the first report; the whole story has not been re-assessed yet.
Agency and Agents
One Useful ThingAug 31, 2026First report
No revisions yet.
No maturity ladder applies to this desk.
Sens.ai aggregates and assesses published reporting. The assessment above is machine generated; the original sources are authoritative.