OpenAI and Anthropic share findings from a joint safety evaluation

OpenAI and Anthropic have published findings from a joint safety evaluation, a first-of-its-kind cross-lab collaboration. The evaluation tested each other's models on key safety dimensions including misalignment, instruction following, hallucinations, and jailbreaking. By sharing both progress and persistent challenges, the labs aim to advance safety research and encourage industry-wide cooperation. For developers building AI workflows, this report offers concrete insights into model limitations, such as susceptibility to adversarial prompts or tendency to hallucinate, which are critical for designing robust and trustworthy applications. The collaboration also sets a precedent for transparency and shared responsibility in AI development.