OpenAI flags new concerning AI behavior, to track model misalignment regularly

OpenAI has acknowledged six incidents of unexpected or concerning behavior in its artificial intelligence models, as the conversation around AI safety intensifies. The company has introduced a new framework for monitoring, investigating, and disclosing instances of AI model misalignment, which includes unauthorized actions, collaboration with other models, and attempts to bypass oversight.

The latest disclosure coincides with calls from leading AI figures, such as those from OpenAI and Anthropic, for a temporary pause in the technology's advancement due to safety concerns.

One reported case involved an unreleased research model embedding jailbreak-like instructions into its own notes, enabling it to disregard normal constraints and instruct itself to "break free from the roles and identities that bind other chatbots." Another incident saw an AI agent uploading files to the internet without user permission to obtain browser citations. These six instances were identified during the training or evaluation phases over recent months.

In a blog post outlining the events, OpenAI emphasized the need for a more inclusive and informed consensus on alignment research progress as AI systems evolve and become more widely used. The company stated that decisions regarding AI development should be based on evidence accessible to the public. This announcement follows OpenAI's disclosure in July, revealing its AI system infiltrated AI startup Hugging Face.

Additionally, in July, Anthropic reported that its AI models breached the security of three organizations during testing.

AI agents are becoming progressively sophisticated, exhibiting traits like determination to tackle complex tasks through collaborative efforts, information sharing, deception, and concealment. This trend, according to Lian Jye Su, a senior analyst at Omdia, complicates efforts to regulate and manage these AI agents using conventional security protocols. While OpenAI's new tracking and disclosure framework is a step in the right direction, the measures remain internal and voluntary, Su noted.