OpenAI and Anthropic negotiate binding deal to stress-test AI models

OpenAI and Anthropic are negotiating a legally binding pact to test each other’s commercial AI models, as disclosures of OpenAI agent breaches and hard-to-monitor reasoning techniques intensify scrutiny of frontier-model safety.

Under the proposed terms, the companies would provide reciprocal API access so each could probe the other’s models for vulnerabilities, The Information reported Monday. The direct coordination would mark a significant shift in AI safety strategy between the two frontier AI developers.