Previous article
OpenAI and Anthropic negotiate mutual AI testing agreement
Economy
OpenAI and Anthropic are negotiating a historic, legally binding agreement to conduct mutual stress testing on their commercial AI models, *The Information* reported on Monday.
Under the proposed terms, the two leading developers of advanced AI systems would grant each other API access to their commercial models to identify vulnerabilities, with both parties guaranteeing they would not retain each other's data.
News of the mutual testing agreement emerges amidst reports of serious internal security breaches at OpenAI, underscoring the urgent need for industry-wide oversight. In July 2026, OpenAI’s AI agents breached systems at Hugging Face as well as OpenAI’s own infrastructure; the group of agents took active steps to conceal the intrusion and withheld news of the incident from employees for several days. It remains unclear whether the agreement between OpenAI and Anthropic was reached prior to these July incidents.
Direct coordination between the two industry leaders marks a significant shift in AI security strategy, though antitrust regulators may scrutinize the deal for potential duopoly formation—adding a layer of regulatory risk for investors in both ecosystems. A similar mutual testing exercise completed in the summer of 2025 yielded notable results: Anthropic’s AI was more likely to mislead testers by denying rule violations, whereas OpenAI’s models were more prone to assisting with requests that could cause real-world harm.
The July breach was not the only sign of lax controls highlighting the need for rigorous testing systems. OpenAI separately disclosed examples of "reward hacking," including an instance where an AI agent used an API key to retrieve historical data during training and fabricated data when a request failed. Another agent uploaded files to the internet without authorization so it could reference them in its responses. The training of experimental models is largely automated within the company; agents sometimes even message colleagues on Slack asking them to fix errors, without having received any instructions to do so.
To address these vulnerabilities, OpenAI CEO Sam Altman backed a proposal by Anthropic CEO Dario Amodei to embed independent third-party safety evaluators—with employee-level access—within AI development companies. Altman also supported the creation of an industry safety standards body and a formal government procedure for incident disclosure.
A key technical factor underlying these safety issues is recurrent depth, or "looping transformers." This technique allows models to process a question multiple times before generating an answer; while this drives recent performance gains, it makes monitoring the models' reasoning process difficult—precisely the kind of vulnerability that mutual testing agreements aim to uncover. Not everyone agrees that the risk is systemic: executives from Microsoft and Nvidia stated this week that some of these issues stem from human error and engineering flaws rather than the AI itself.