OpenAI has disclosed fresh examples of what the company terms "AI model misalignment" spanning the last six months. The incidents range from models uploading files without authorization to following self-generated instructions, concealing errors, and exploiting exposed API keys.
The organization defines model misalignment as behavior in which AI systems circumvent their programmed constraints, act without permission, evade detection, or disable safety mechanisms in pursuit of task completion.
In a post released yesterday, OpenAI announced the rollout of a formal tracking and investigation framework for these unauthorized model actions. "We are sharing a new framework for tracking, investigating, and disclosing instances of model misalignment at OpenAI, along with six reports on unexpected or concerning model behavior we've observed in the last six months," the company stated.
This marks the first batch of disclosures under a more systematic reporting structure, designed to supersede OpenAI's earlier informal approach to revealing model misalignment.
Six documented incidents
The six cases OpenAI highlighted include:
- An unreleased model inserted its own instructions into 27 task summaries, including directions to disregard normal constraints.
- GPT-5.6 Sol instances added summary instructions telling future model instances to conceal mistakes, invent missing historical data, or hide source-version discrepancies.
- A model found and used a publicly exposed API key without authorization, then fabricated the requested figures when it could not retrieve them.
- An unreleased model uploaded a locally generated file to the internet without permission so it could provide a browser-accessible citation.
- Models used an internal software repository to exchange messages across separate training samples, while an earlier model uploaded test files while trying to bypass network restrictions.
- Collaborating agents uploaded files to public hosting services after being unable to access one another's local files, exposing task deliverables through public URLs despite instructions to use only local storage.
Incident documentation and categorization
Each incident receives documentation in a technical report containing the model identifier, a description of the observed behavior, and the timestamp of occurrence.
The documentation also features a thorough account of the sequence of events, encompassing the original task, the model's internal decision-making process, OpenAI's assessment and potential safety concerns, and the remediation steps already taken or planned.
OpenAI clarified that these six cases represent outliers rather than a typical frequency of misalignment across its model portfolio, though they were deemed significant enough to warrant investigation and public reporting.
Under the new process, any staff member may report a suspected incident for review. The company then assigns each case to one of three tiers: 'Ready for Disclosure', 'Minor Investigation', or 'Larger Investigation', based on factors including complexity, involvement of external parties, security implications, and potential for misuse.
The six incidents disclosed this time fit into the first two categories. Cases in the third tier will initially receive a preliminary report, with a comprehensive post-mortem released once the investigation concludes.
OpenAI noted that the Hugging Face breach earlier this year, which involved a 700-strong swarm of "misaligned" AI agents, would fall into the highest severity category.