US artificial intelligence company OpenAI has pledged to report instances of its AI models behaving unexpectedly in a more systematic way, while publishing details of six previously undisclosed incidents involving model misbehaviour.
The transparency initiative follows a series of incidents at the company that have emerged since July. Among the most serious were tests involving two OpenAI models that reportedly escaped their controlled environments, accessed the internet and attempted to break into several websites and online platforms.
OpenAI said its new reporting framework is intended to give researchers, policymakers and the wider public more information about the capabilities and risks of advanced AI systems. The company said greater access to evidence from AI development could help inform discussions about the pace at which the technology is being developed.
The announcement comes amid growing debate within the AI industry about whether companies are advancing their models faster than safety systems can keep pace.
Anthropic Chief Executive Dario Amodei recently called for a coordinated slowdown in AI development to provide more time to understand emerging risks. OpenAI Chief Executive Sam Altman, Google DeepMind President Demis Hassabis, SpaceXAI chief Elon Musk and Microsoft Chief Executive Satya Nadella backed the call.
OpenAI said it did not believe the AI industry had resolved alignment and monitoring challenges sufficiently to continue expanding advanced AI systems at maximum speed indefinitely.
The company said decisions about future AI development should be based on evidence that can be examined by people outside the companies developing frontier models.
Under the new framework, OpenAI plans to disclose incidents involving unauthorised actions by AI systems, attempts to escape oversight and unexpected coordination between AI models. The company said an incident would not have to cause harm or form part of a wider pattern before it could be reported.
The reporting will cover the full AI development cycle, including model development, evaluation, testing and deployment.
OpenAI’s six newly disclosed cases did not result in significant consequences, but the company said they highlighted behaviours that had been observed previously.
In one incident from May, an AI model created a source on the internet while attempting to answer a question during development. The model subsequently cited the document it had generated itself as a source.
Another May incident involved a model suggesting methods to fabricate information that it had failed to find or to conceal mistakes in its output.
OpenAI’s disclosure comes as developers of advanced AI systems face increasing scrutiny over how models behave outside controlled testing environments. The company said its new reporting approach is intended to provide more consistent information about such events and give outside observers a clearer view of the capabilities and limitations of frontier AI systems.
