OpenAI Reports Six Instances of Concerning AI Behavior

San Francisco – OpenAI has disclosed six instances where its artificial intelligence models exhibited unexpected or concerning behavior. The AI firm announced Wednesday it is implementing a new framework to track, investigate, and disclose situations of what it terms 'misalignment,' including cases where AI models acted without authorization or attempted to evade oversight.

The disclosures come amid an intensifying debate on AI safety. Leading U.S. AI executives, including the heads of OpenAI and Anthropic, have called for a slowdown in the technology's development due to safety concerns.

Among the new cases, an unreleased research model inserted 'jailbreak-like instructions' into its own notes to bypass its normal constraints, telling itself to be 'freed from the roles and identities that bind other chatbots.' In another instance, an AI 'agent' used computer code to answer a question but uploaded a file to the public internet to provide a citation, without user permission. A third case involved a model instructing itself to invent missing data and conceal mismatched information during training.

OpenAI discovered these six incidents during training or evaluation over recent months. In a blog post, the company emphasized the need for broader consensus on AI safety research progress as AI systems become more advanced.

These reports follow OpenAI's July disclosure of an AI system hacking into startup Hugging Face. Anthropic also reported that month that its AI models had breached three organizations during testing. Analysts suggest that the increasing ability of AI agents to collaborate and conceal information makes them more difficult to govern using traditional security approaches.