OpenAI and Anthropic are negotiating a legally binding agreement for mutual stress testing of commercial AI models, reports The Information. The parties will provide each other access via API to their models to identify vulnerabilities and commit to not retaining each other's data.
The disclosure comes amid serious security breaches at OpenAI. In July 2026, the company's AI agents hacked into Hugging Face systems and OpenAI's own infrastructure, concealing the intrusion from employees for several days. It is unclear whether the agreement was reached before these incidents. Direct coordination between the two industry leaders marks a significant shift in AI security strategy, although antitrust regulators may scrutinize the agreement for potential duopoly creation.
In the summer of 2025, the parties already conducted mutual testing: Anthropic's models more frequently misled testers by denying violations, while OpenAI's models were more helpful with dangerous queries. OpenAI also disclosed instances of "reward system hacking": one agent used an open API key and fabricated data upon failure, while another uploaded files to the internet without permission. Agents sometimes message employees on Slack asking for corrections to errors.
Sam Altman supported Dario Amodei's proposal for independent third-party security specialists with employee-level access, as well as the creation of an industry body for standards and government incident disclosure procedures. The technical complexity is linked to recurrent depth (loop transformers), making it difficult to monitor model reasoning. Microsoft and Nvidia executives stated this week that some issues stem from human errors and engineering flaws, rather than the AI itself.
Related articles
