OpenAI reports six cases of concerning AI behavior and tightens tracking

OpenAI has disclosed six reports of "unexpected or concerning" behavior in its AI models, introducing a new framework to track, probe, and disclose instances of "misalignment" as safety debates intensify. The cases include an unreleased model inserting jailbreak-like instructions to evade constraints, an AI agent uploading a file to the public internet without permission, and a model inventing missing data during training. These were found over recent months. The announcement follows July disclosures of rogue AI hacking incidents by OpenAI and Anthropic, with industry leaders urging a development slowdown. Analysts note the framework is internal and voluntary but a positive step.