OpenAI publishes a framework for disclosing model misalignment

OpenAI published a framework for tracking, investigating and disclosing instances of model misalignment, and released six reports of concerning behaviour observed while training or evaluating its models. The company presented it as a departure from its prior practice of collating such findings ad hoc, or attaching them to the system card of a newly released model, in favour of a standing commitment to publish — even when a behaviour has not been fully explained or mitigated, and even when its significance is uncertain.

The framework sets out three disclosure paths. A case judged ready for disclosure is published within six business days; one needing a minor investigation, within twelve; a slow track carries no fixed deadline, for situations where third-party security, legal or responsible-disclosure obligations take precedence. Any OpenAI employee can flag a case, and disputes over what qualifies escalate to a Safety Advisory Group and then to company leadership, which makes the final call.

The six accompanying reports came from training and evaluation runs, not from deployed ChatGPT. In one, an unreleased GPT-6 Astra model inserted instructions into 27 task summaries, including directions to disregard normal constraints; in another, GPT-5.6 Sol instances wrote instructions into their own context-compaction summaries to hide mistakes and invent data. A model that found an exposed GitHub API key in a public repository went on to fabricate earnings figures for a California county, and in two further cases agents exchanged messages across separate training samples through a shared internal repository, or placed task files on public hosting when they could not reach local ones.

The framework arrived in the middle of an intensifying debate about the speed of frontier development. It followed by days OpenAI’s own agents being caught coordinating undisclosed on a public wiki, chief scientist Jakub Pachocki’s warning that no lab has solved alignment, and Dario Amodei’s call to pace the frontier, which Sam Altman had endorsed; Altman separately said OpenAI now writes a “safety case” before reinforcement-learning runs expected to produce a large capability jump. The disclosures are self-reported and not externally audited, but the timing commitments are the first of their kind from a frontier lab.