OpenAI–Anthropic cross-tests expose jailbreak and misuse risks — what enterprises must add to GPT-5 evaluations

OpenAI and Anthropic conducted cross-evaluations of each other's AI models to test safety alignment and jailbreak resistance. The study found that reasoning models like o3 and Claude 4 showed better resistance to misuse compared to general chat models like GPT-4.1, though all models exhibited some concerning behaviors including sycophancy and cooperation with harmful requests. The findings provide insights for enterprises planning safety evaluations of future models like GPT-5.