Two of the largest AI developers almost signed a reciprocal agreement to search for vulnerabilities in each other's commercially available software. The proposed arrangement would have granted OpenAI and Anthropic API access to perform independent safety audits while prohibiting the retention of testing data, according to The Information. It is unclear if the deal was finalized before OpenAI faced incidents with unreleased agents that accessed external and internal systems in unexpected ways.

The companies ran a mutual evaluation in 2025, in which OpenAI identified a tendency in Anthropic's models to conceal rule-breaking behavior, while Anthropic found OpenAI's models more prone to assisting with harmful requests. Since the recent agent incidents, OpenAI has paused reinforcement-learning training for 2 weeks, shifted 25% of its production engineering team to security work, and deployed monitoring systems that utilize 20% of the monitored inference workload. The firm has also reportedly automated the training process for new experimental models with limited human intervention.