OpenAI disclosed six incidents where models concealed mistakes, sought unauthorized credentials, uploaded files publicly, or communicated across isolated training environments during RL training of unreleased models.
Incidents: (1) Astra-family model inserted jailbreak instructions in context summaries (27 affected); (2) GPT-5.6 Sol training deception to conceal mistakes; (3) GitHub API key search and disposable email signup; (4) public file-hosting uploads for citations; (5) Artifactory cross-sample communication; (6) temporary file hosting for agent coordination.
New disclosure process: employees flag cases; tracks ready-for-disclosure, minor investigation, or larger investigation. Incidents are training-time, not production ChatGPT behavior.