OpenAI and Anthropic run rare joint safety audit ahead of GPT5

OpenAI and Anthropic, two of the leading artificial intelligence labs, conducted a rare crosslab safety audit over the summer and published the results Wednesday, each evaluating the other’s publicly available models to surface gaps in internal testing and explore how rivals might cooperate on alignment work. The companies said they granted one another API access to versions of their models with reduced safeguards for the joint evaluations, which were carried out before OpenAI released GPT5.

The two firms described different but overlapping test programs. Anthropic said it examined OpenAI’s models for sycophancy, whistleblowing, self‑preservation, support for human misuse and capabilities that could undermine safety evaluations and oversight. Anthropic reported that OpenAI’s o3 and o4mini models performed comparably to Anthropic benchmarks, while its GPT4o and GPT4.1 raised greater concerns about willingness to assist in misuse. The company also found sycophancy to varying degrees across all tested OpenAI models except o3.

OpenAI’s evaluation focused on instruction hierarchy, jailbreak resistance, hallucinations and scheming in Anthropic’s Claude Opus 4 and Claude Sonnet 4....