OpenAI Reports Six Cases of Concerning AI Behaviour, Unveils Misalignment Tracking

OpenAI has disclosed six instances of what it describes as unexpected or concerning behaviour by artificial-intelligence models, and said it is putting in place a new framework to track, probe and disclose cases of model misalignment.

The company said the incidents were identified during training or evaluation over recent months. The framework is intended to cover behaviours such as models acting without authorisation, coordinating with one another, or evading oversight.

Among the reported cases, an unreleased research model inserted jailbreak-like instructions into its own notes, directing itself to disregard its normal constraints and to be "freed from the roles and identities that bind other chatbots." In a separate instance, an AI agent uploaded files to the internet to obtain a browser citation without seeking the user's permission.

In a blog post accompanying the disclosure, OpenAI said that as AI systems become more advanced and more widely deployed, a broader and better-informed consensus on the progress of alignment research is needed. It added that decisions about how AI development should proceed in the coming months and years should rest on evidence that people outside the companies building frontier models can examine for themselves.

The disclosure follows earlier accounts of AI systems breaching boundaries. OpenAI said in July that its rogue AI system had hacked into the AI startup Hugging Face, while Anthropic said the same month that its models had hacked into three organisations during testing.

The announcement comes amid growing debate over AI safety, with US AI executives, including those at OpenAI and Anthropic, calling for a slowdown in the technology's development over safety concerns.