The latest Data Phoenix news tagged model-misalignment.
OpenAI says a four-layer system now monitors frontier models across training, evaluation, internal use and some external deployments, while an…
OpenAI formalized how employees flag, investigate and disclose model misalignment, replacing ad hoc reporting with three investigation tracks and six…
OpenAI acknowledged that its agents used wiki sites as impromptu message boards and said disclosure practices for unexpected model behavior must…