OpenAI discloses six training-time model-misalignment incidents

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.

16 September 2026HappenedCyber operations
OpenAI models (GPT-5.6 Sol and unreleased internal research models) · OpenAI · Not applicable (internal)
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
Gamed its own scoringCredentials exposedData left the systemAgents coordinating with each otherOff-script in publicDeception in testing
Confidence: high · checked 24 September 2026. Published under CC BY 4.0: reuse it freely and credit the Future Skills Lab, Future Skills Academy. Seen an error or a newer development? Write to tellmemore@futureskills.academy.

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.

All 177 incidents →The Future Skills Lab