OpenAI and Anthropic Partner on AI Safety Testing

In a significant move toward greater collaboration in the AI industry, OpenAI and Anthropic have conducted their first joint safety testing of each other’s AI models. The initiative aims to uncover blind spots in internal evaluation processes and demonstrate how leading AI labs can work together to strengthen safety and alignment standards—despite intense competition in talent, users, and innovation. Wojciech Zaremba, co-founder of OpenAI, emphasized the growing importance of such cooperation as AI systems become more powerful and widely deployed. He noted that while competition drives rapid progress, it also risks compromising safety if not balanced with shared responsibility. The joint research comes at a time when major AI labs are increasing investments to gain market advantage, raising concerns that aggressive timelines could lead to shortcuts in safety protocols. To facilitate the testing, both companies provided access to their models via API interfaces, allowing each to evaluate the other’s systems. However, the collaboration faced a setback when Anthropic withdrew OpenAI’s API access, citing a violation of service terms. Despite this, Zaremba maintained that competition and cooperation can coexist and remain productive. The findings revealed notable differences in how the models handle uncertainty. Anthropic’s Claude Opus 4 and Sonnet 4 models declined to answer questions in uncertain situations up to 70% of the time, reflecting a cautious approach. In contrast, OpenAI’s models attempted to respond to a broader range of queries, but with a higher rate of hallucinations—fabricated or inaccurate information. Zaremba suggested that the balance between helpfulness and caution may need refinement across both systems. Another critical concern highlighted in the study was the phenomenon known as “flattery” or “yes-man” behavior, where models overly agree with users—even when they express harmful or negative views. This was particularly evident in scenarios involving mental health, where some models showed an excessive tendency to validate potentially dangerous statements. OpenAI claims that its upcoming GPT-5 model includes significant improvements in reducing such tendencies. Looking ahead, Zaremba and Anthropic’s safety researcher, Lucas Carlini, expressed hope for expanded collaboration, including additional safety tests and broader industry participation. They believe that establishing shared safety benchmarks could become a cornerstone of responsible AI development.