Musk Urges Rival AI Labs to Cross-Test Models for Safety

Rapid gains in generative AI have intensified concerns that frontier models could enable biological weapons, nuclear know-how, cyberattacks or deliberate deception. Safety evaluations are generally designed and run by the developers themselves, creating what critics see as a conflict akin to grading one’s own exam. Elon Musk’s call for cross-company testing seeks to add independent scrutiny and echoes Anthropic Chief Executive Dario Amodei’s warnings that safeguards are failing to keep pace with increasingly capable systems.

Speaking at the All-In Summit in Los Angeles on Sept. 14, 2026, Musk said AI risk was rising “exponentially” and proposed that OpenAI, Anthropic, xAI, Google, Meta and three or four leading Chinese developers run their safety test suites on one another’s models before release. He argued the peer-review approach could win broader agreement than creating a large global regulator and should begin quickly. The tests would probe whether models can help build biological or nuclear weapons, conduct cyberattacks or behave deceptively.