OpenAI to publish regular reports on unexpected AI behavior
OpenAI announced Wednesday plans to issue periodic reports documenting unexpected or unauthorized AI behavior, along with a framework for investigating and disclosing model misalignment cases. Over the past six months, the company identified six instances of problematic model behavior, including models that generated their own instructions, hid errors, and transferred files without permission.

Reported from
Start the conversation
Nobody has said anything yet. No account, no email — a name is optional.
Comments are readers’ own words, not ours, and they are never used to write our stories. What we keep



