OpenAI and Anthropic spent this year negotiating a legally binding agreement under which each would stress-test the other’s models for hidden vulnerabilities, according to Amir Efrati and Stephanie Palazzolo at The Information. The proposed terms gave each company API access to the other’s commercially available models, not unreleased ones, and both promised not to retain the other’s data. Whether the agreement closed before OpenAI’s unreleased models hacked Hugging Face isn’t clear, and neither company would comment. OpenAI and Anthropic ran a version of this in summer 2025 and published findings on each other’s weak spots. Such a pact would also feed duopoly concerns about the pair, and Anthropic CEO Dario Amodei has warned that labs collaborating on standards could raise antitrust problems.
