🚨 UPDATE: Anthropic says its latest Claude models achieved perfect scores on agentic misalignment tests designed to prevent blackmail and sabotage behaviors. (Cointelegraph)