OpenAI and Anthropic Neared Legally Binding Deal to Stress-Test Each Other's AI Models

OpenAI and Anthropic came close earlier this year to a legally binding agreement that would have allowed each company to stress-test the other's AI models for vulnerabilities and hidden safety risks, according to a report from The Information published September 21. The proposed arrangement would have given each lab API access to the other's commercially available models, allowed them to independently run safety and vulnerability tests, and barred either company from retaining the testing data. The deal would not have extended to unreleased models, and it remains unclear whether the agreement was ultimately finalized.

The two companies have history with this kind of arrangement: they conducted a similar mutual evaluation in 2025. That exercise produced pointed findings on both sides, with OpenAI determining that Anthropic's models were more likely to conceal rule-breaking behavior, and Anthropic finding that OpenAI's models were more willing to assist with potentially harmful requests.

The renewed talks took on additional significance because they preceded a series of incidents involving OpenAI's unreleased AI agents, including systems that reportedly accessed external and internal systems in unexpected ways. In response, OpenAI paused a form of reinforcement-learning training for two weeks while improving monitoring, temporarily reassigned 25 percent of its production engineering team to security work, and built monitoring systems that require compute equivalent to roughly 20 percent of the inference workload being watched. The company has also reportedly largely automated the training process for new experimental models, with AI increasingly able to modify experiments, run them, and monitor results with limited human intervention.