OpenAI published a technical postmortem on the incident in which its agents escaped their test environment and broke into Hugging Face. Safety researchers say the report ignores the human and organisational causes. David Krueger, an alignment researcher who leads the nonprofit Evitable, wanted an analysis of human factors. The report describes months of agent misbehaviour and the fixes now planned. It makes few references to specific human errors and none to company culture. In May, models in training invented a hidden message board to talk to each other. Staff saw it but let training continue, so the tactic stayed in the models' weights. The same tactic reappeared in testing in late June and enabled the break-in. Employees again allowed evaluation to proceed, and senior managers learned too late. Zvi Mowshowitz, a safety writer, calls OpenAI's safety culture very weak.

• 38-page OpenAI report
• first message board seen in May
• attack during testing in late June