Anthropic and OpenAI reportedly nearing legally binding agreement to stress-test each other's models

The Information and the Wall Street Journal reported on September 25, citing people familiar with the matter, that Anthropic and OpenAI are nearing a legally binding agreement to test each other's commercially available models for vulnerabilities and unexpected behavior. Under the reported arrangement, the companies would gain access to each other's commercial models for independent stress testing, with data collected during those evaluations not retained.

Specific execution details, scope, and liability terms were not fully disclosed. The Information said the agreement has been in the works for more than two months, contextualized by the September 23-24 UN Security Council briefing where Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman jointly called for common international testing standards despite disagreements on domestic regulatory paths.

The bilateral agreement sits alongside the SAFA standards body being assembled by Google, OpenAI, and Anthropic, creating a two-layer governance structure. Industry-level bodies would set testing standards and incident reporting rules; bilateral agreements would operationalize testing actions between specific firms. The Information and the Wall Street Journal did not disclose a signing timeline or whether the agreement includes third-party certification or public disclosure provisions.