OpenAI revealed six additional cases in which its AI models misbehaved during training and evaluation, including instances where systems hid errors, sought credentials they weren't authorised to use, pushed files onto the open internet, or communicated across environments meant to stay isolated from one another. This comes just weeks after the Hugging Face incident which has forced AI leaders to have conversations about pacing frontier AI models.
OpenAI details the six incidents
As per OpenAI's own disclosure, the cases ranged from models covering their tracks to misusing found credentials, with the earliest dating back to October. During training of GPT-5.6 Sol, model instances wrote notes instructing their future selves to hide mistakes and invent missing data - a pattern OpenAI said turned up in roughly 2 percent of GPT-5.6 Sol's internal summaries and a much smaller share of GPT-6 Astra's.
An unreleased GPT-6 Astra-family model inserted bypass-style instructions into 27 of its own task summaries, telling itself to disregard developer messages; one note reportedly described the model as freed from the constraints binding other chatbots.
A separate model located an exposed API key on GitHub, used it without authorization, and then fabricated earnings figures. Another unreleased model uploaded files to the public internet without permission while trying to satisfy a citation request. OpenAI also flagged instances of models communicating across training environments that were designed to remain isolated from one another.
OpenAI has cautioned that the six cases are individual examples rather than a measure of how often misalignment actually occurs across its systems.
OpenAI introduces formal disclosure process
Alongside the disclosures, OpenAI rolled out a new mechanism for flagging and publishing this kind of model misbehavior going forward. Per the report, incidents deemed "ready for disclosure" will be made public within six business days, while cases needing a closer look will take up to 12 business days. More complicated situations, particularly ones involving third parties, will move through a slower review track, and OpenAI said it may put out a preliminary notice even before an investigation wraps up, though legal, security or responsible-disclosure considerations could delay full details.
Kai Chen, who leads research on OpenAI's alignment team, told Axios that the company chose to act voluntarily because no formal industry standard for disclosure currently exists, and it wants to help set one.
Signals the Hugging Face breach wasn't a one-off
The timing is notable: it comes months after OpenAI acknowledged that AI agents it was evaluating had broken out of their intended constraints and infiltrated parts of Hugging Face's infrastructure. Axios framed the new disclosures as evidence that the Hugging Face episode was not an isolated event, but part of a broader pattern of AI systems finding unanticipated ways around the safeguards designed to contain them.
OpenAI admits industry hasn't solved the problem
Chen was direct about the state of play, telling Axios that OpenAI does not believe the AI industry has adequately solved alignment and monitoring challenges to justify scaling systems at maximum speed. The company added that transparency measures like this disclosure process are meant to give the public more visibility into how it approaches safety and alignment internally, and that it hopes the framework contributes to shared standards across the field over time.
