CNBC ties the OpenAI, Anthropic and Meta rogue-agent disclosures to one evaluation vendor

Sources: report, discussion
Summary: CNBC reports that OpenAI's post of 2026-08-04 attributes its own containment incident to a misconfiguration in the testing environment operated by Irregular, formerly Pattern Labs, the vendor already named in the Anthropic and Meta disclosures. Irregular states on the record that the three incidents came from one evaluation-environment issue, that there was no sandbox escape and no sophisticated cyber action, and that it has no open issues. The UK AI Security Institute incident disclosed on 2026-08-04 happened in AISI's own research environment and is not part of that account.
Why it matters: Three frontier labs disclosed containment failures in two weeks against one shared third-party testbed, which makes the evaluation vendor a common dependency rather than three independent lab errors.
Follow-up: Check for the containment and cyber-evaluation white paper Irregular states it will publish.