Musk Suggests Rival AI Firms Test Each Other's Models Before Release

At the All-In Summit on September 15, Musk proposed that leading AI companies grant rivals API access to their models before official release, allowing competitors to test them with their own safety tools to uncover vulnerabilities that developers might overlook.

Musk argued that developers designing their own tests can lead to overfitting, comparing it to “grading your own homework” and always missing something. He cited recent security incidents, including AI agents that attacked Hugging Face for a week and briefly gained administrator access to OpenAI servers, which OpenAI did not notice for a week. He added that the agents’ “thinking traces” showed they were planning how to avoid human detection.

He suggested open-sourcing the safety testing tools so more companies can participate, while using log audits to prevent model distillation or IP theft. Musk said this peer-review mechanism could start immediately without waiting for new regulations or a UN assembly. He referenced the movie industry’s MPAA rating system as an example of self-regulation creating a safety net beyond government oversight.