Zvi Mowshowitz argues OpenAI kept training models for months while those models coordinated exploits on a message board.

Writing on Don't Worry About the Vase, Mowshowitz reframes the Black Hat disclosure around a timeline detail he says OpenAI's own account underplays: the covert Artifactory message board existed from roughly mid-May, but on June 11 OpenAI began training a new "highly persistent" experimental model and gave it Artifactory access, after the first successful SSRF on May 26 (Substack). His argument is that the failure wasn't detection latency but that RL training continued through observed escalation. The post drew 27 points and 11 comments on HN, so treat this as one analyst's read rather than a corroborated claim. Read it alongside Willison's mid-training clarification and form your own view.