OpenAI and Anthropic previously negotiated a legally binding agreement to conduct mutual stress testing of their large language models. The discussions took place earlier this year, involving legal teams from both companies to define the scope of vulnerability assessments.
The proposed deal stipulated that both companies would only gain access to commercialized model APIs, with no access to unreleased models. Furthermore, both parties committed to not retaining any data obtained during the testing process.
This approach aligns with a concept suggested by Elon Musk, who proposed that competing AI labs perform peer reviews of model security before commercial release. Sam Altman, CEO of OpenAI, has also expressed support for independent third-party assessments and industry-wide safety standards.
OpenAI is currently exploring various collaboration models with other firms and government entities to improve safety. These discussions now cover security mechanisms for both pre-training and pre-release stages, following internal concerns regarding existing security measures.
Some industry players, including executives from Microsoft and Nvidia, argue that current AI risks stem from human error rather than the technology itself. Additionally, there are concerns that such partnerships between industry leaders could lead to antitrust scrutiny.
