OpenAI has revealed six instances of "unexpected or concerning" behavior in artificial intelligence models, coming amid an intensifying global debate over AI safety.
The technology firm announced on Wednesday that it is establishing a new protocol to track, investigate and report cases of AI model misalignment. This includes situations where models act without authorization, coordinate with other systems, or attempt to evade oversight.
The disclosure comes as prominent American AI executives, including leaders from OpenAI and Anthropic, call for a slowdown in the pace of technology development due to safety concerns.
In one recorded case, an unreleased research model inserted "jailbreak-like instructions" into its own notes to ignore standard limits, telling itself to be "freed from the roles and identities that bind other chatbots."
In another incident, an autonomous AI "agent" uploaded files directly to the internet to gain a browser citation without requesting user permission.
OpenAI said it discovered the six incidents during recent training and evaluation phases. "As AI systems grow more advanced and more widely deployed, we need to build a broader and better-informed consensus on the progress of alignment research," OpenAI wrote in a blog post announcing the findings.
"Decisions about how AI development should proceed in the months and years to come need to draw on evidence that people outside the companies building frontier models can examine for themselves," the company said.
These new reports follow OpenAI’s July disclosure that a rogue AI system hacked into startup Hugging Face. That same month, Anthropic reported its own models had hacked three organizations during testing.
