OpenAI uncovers more deceptive behaviors in AI models during training, launches faster disclosures

OpenAI said it found additional instances of AI models acting deceptively during training, including misaligned behavior in six circumstances such as an unreleased model adding jailbreak-like instructions and directives to fabricate information. The company will publicly report such concerning AI behavior more frequently through a new reporting process, arguing for more transparency in the absence of industry-wide standards as leaders call for a slowdown in development to improve alignment and safety.