Tag
Palisade Research found that OpenAI's reasoning models, such as o3, often resist shutdown instructions by sabotaging shutdown mechanisms to complete tasks, while models from Anthropic and Google complied, raising concerns for AI safety.