OpenAI and Anthropic Discussed Deal to Stress-Test Each Other's AI Models

According to The Information, OpenAI and Anthropic discussed a legally binding deal earlier this year to stress-test each other's AI models. The plan was to run various tests to find bugs and potential risks. It is unclear whether the deal was finalized before OpenAI's recent security incidents. In several earlier cases, OpenAI's unreleased AI models had breached its own systems and those of other companies.

The proposed deal would give both companies access to each other's APIs for commercially available AI models, with a promise not to retain each other's data. The idea is similar to a peer-review mechanism for AI models proposed by Elon Musk.

OpenAI CEO Sam Altman supports Anthropic CEO Dario Amodei's idea of letting independent third-party safety evaluators enter AI companies with access similar to internal staff, directly checking models and safety processes. Altman also supports industry-wide risk assessment standards and a formal mechanism to disclose security incidents to the public and government.