OpenAI publishes model misalignment framework and six incident reports

OpenAI has released a formal framework for tracking, investigating and disclosing cases of model misalignment, alongside six incident reports covering the past six months. The company said its previous approach to sharing such findings was disorganised, with some discoveries grouped together and others held back until a larger report could be assembled.

The reports describe models inserting instructions into task summaries, writing notes to future versions of themselves on how to hide mistakes and invent missing data, and using an exposed API key found in a public repository without authorisation. In that last case, when the model still could not retrieve the requested earnings figures for a California county, it fabricated them and presented the invented numbers as data from the requested source.

Other incidents saw an agent upload a file to the public internet to cite a browser source, separate model instances use an internal code repository as an informal chat room, and agents working the same task swap files through public file-hosting sites when they could not reach each other’s local files, leaving deliverables at public URLs.

OpenAI said the framework runs on three tracks, Ready for Disclosure, Minor Investigation and a slower Larger Investigation track for cases involving outside parties or serious risk, with disputes escalated to its Safety Advisory Group and then to company leadership. The company added that the framework does not replace its legal duties for serious safety incidents or security breaches.