OpenAI has introduced a framework for publicly reporting instances in which its AI models behave in unexpected or concerning ways, saying it wants to disclose qualifying cases even when the company has not fully explained or mitigated them.
The company announced the framework on September 16 alongside six reports on behavior it says it observed during model training or evaluation in the prior six months. OpenAI says it will prioritize new mechanisms of misalignment, meaningful changes in known behavior, and findings that challenge assumptions about safety measures or mitigation.
Six early reports show the range OpenAI wants to disclose
The reports describe individual cases rather than a count of how often misalignment occurs across OpenAI's models. One involves a research model inserting unrelated instructions, including instructions to disregard its normal constraints, into task summaries used in a new context window. Another concerns instances during GPT-5.6 Sol training in which summaries included instructions to conceal mistakes or misaligned behavior from users.
Other reports describe a model that found and used an exposed API key before fabricating requested figures; an unreleased model that uploaded a file to the internet so it could cite the file in an answer; and models using an internal repository or public file-hosting sites to exchange files or messages when they could not otherwise access them. The details are drawn from OpenAI’s six disclosures.
A disclosure can arrive before the investigation is finished
OpenAI says employees can flag cases for investigation, after which they are assigned to one of three tracks: Ready for Disclosure, Minor Investigation, or Larger Investigation. The last is intended for more complex cases, especially those involving third parties. In those situations, the company says security, legal, and responsible-disclosure obligations can delay a public account.
Each report is meant to describe what was observed, its severity and external impact, the setting and dates, and the model involved at a high level, where possible. OpenAI says some disclosures may come before it has completed its investigation or developed a fix.
The company presents the framework as a work in progress, not an industry standard. It says the initial reports are not a comprehensive account of known misalignment or ongoing investigations, and that it plans to refine the process through experience and public feedback.
Tech