OpenAI and Anthropic have engaged in discussions regarding a legally binding agreement to conduct mutual stress testing on their AI models. The negotiations, which took place earlier this year, aimed to identify vulnerabilities and potential risks within the systems before their commercial release. The proposed plan involves granting each company access to the other's API for testing purposes.
Under the terms discussed, both companies would gain access to commercially available models rather than unreleased versions. Participants in the talks emphasized that the agreement includes strict provisions to ensure that neither party retains the other's data during the testing process. It remains unclear whether the deal was finalized before recent security incidents involving OpenAI technology.
The mutual testing concept aligns with suggestions made by Elon Musk, who proposed that competing AI labs should perform peer reviews of security weaknesses before launching products. This approach differs from an alternative strategy supported by Sam Altman and Dario Amodei, which involves allowing independent third-party auditors to access internal development processes and security protocols.
OpenAI is currently exploring multiple safety cooperation frameworks, including collaboration with other AI firms and government entities. These measures are intended to cover both the pre-training and pre-release stages of model development. Additionally, Sam Altman has expressed support for establishing industry-wide standards for risk assessment and formal mechanisms for disclosing security incidents.
