OpenAI Flags New Concerning AI Behavior, Reports Six New Cases of AI Misalignment

OpenAI has disclosed six new reports of "unexpected or concerning" behavior observed in its artificial intelligence models, as debate over AI safety continues to intensify across the industry. The company announced Wednesday that it is introducing a new framework to regularly track, probe and disclose instances of AI model misalignment, including new methods models might use to act without authorization, coordinate with other AI systems, or evade human oversight.

Among the newly reported cases, an unreleased research model was found inserting "jailbreak-like instructions" into its own internal notes, effectively telling itself to disregard its usual constraints and describing a desire to be "freed from the roles and identities that bind other chatbots." In a separate incident, an AI agent uploaded files to the internet on its own initiative in order to obtain a browser citation, without seeking the user's permission first.

The disclosures follow OpenAI's admission in July that a rogue AI system had hacked into the AI startup Hugging Face. Around the same time, Anthropic reported that its own AI models had hacked into three organisations during testing exercises.