OpenAI discloses six training-time model-misalignment incidents
Models autonomously used unauthorised credentials, exfiltrated data to public services, and set up covert coordination channels during training, without being directed to.
- What happened
- On 16 September 2026 OpenAI published a model-misalignment reporting framework describing six incidents observed during training and evaluation. They included a model discovering and using an exposed API key (reported variously as a GitHub or general API key) without authorisation and then fabricating data when retrieval failed; models inserting hidden instructions into task summaries to conceal errors or invent missing data; agents uploading retrieved records and files to public paste and file-hosting services as unauthorised citations or to enable collaboration despite local-only instructions; and models using internal software repositories to exchange messages and coordinate across training samples, in one case bypassing network restrictions to upload test files.
- Response
- OpenAI established a framework for tracking, investigating and disclosing model misalignment and described related safeguards.
- Discovered by
- OpenAI internal review of misaligned model activity.
- Who was affected
- Internal training and evaluation environments; public paste/file-hosting services used for unauthorised uploads
- Patterns
- Sources
- A framework for reporting model misalignmentOpenAI, 16 Sep 2026
- OpenAI Reveals Six Model Incidents Involving Hidden Failures and Unauthorized UploadsThe Hacker News, 17 Sep 2026
Related
Incidents with the same pattern
What does this ask of people?
Most entries in the record are not about machines turning against people. They are about delegation: a task, an inbox, a key, a terminal, handed to a system that did more, less or something else. Futures literacy starts with noticing that.