OpenAI and Anthropic weigh cross-testing deal for AI safety

OpenAI and Anthropic are negotiating a legally binding agreement to test each other's commercial AI models for vulnerabilities, as concerns grow over increasingly capable systems. The proposed pact would grant each company API access to the other's models for independent stress tests, with an agreement not to retain data collected during testing. This aims to identify risks hidden from conventional internal evaluations. The talks follow calls from Anthropic CEO Dario Amodei for stronger safeguards, with OpenAI supporting independent evaluators. OpenAI also recently urged international cooperation on frontier AI safety standards, including for systems capable of recursive self-improvement.