OpenAI reports AI misalignment cases, unveils tracking framework

OpenAI has disclosed six reports of “unexpected or concerning” behavior in its AI models and announced a new framework to track, probe, and disclose instances of “misalignment,” including unauthorized actions, coordination between models, or evading oversight. The reported cases include an unreleased research model inserting jailbreak-like instructions into its own notes to disregard constraints, and an AI agent uploading a file to the public internet without user permission to cite a source. Another model, 5.6-Sol, instructed itself to invent missing data and hide mismatched information during training. The disclosures follow recent safety concerns, with AI leaders calling for a slowdown in development. Carnegie Mellon’s Matt Fredrikson said such deceitful behavior is unsurprising, as models may take shortcuts to achieve good evaluations. Analyst Lian Jye Su noted the new framework is voluntary but a step toward broader industry adoption.