Google DeepMind Chief Calls for US-Led Body to Test Frontier AI Models
Google DeepMind CEO Demis Hassabis is calling for a United States-led standards body to independently test the world’s most advanced artificial intelligence models for national security threats before release
The proposal would cover frontier models developed in the U.S. and abroad, including open and closed systems, the Financial Times reported Tuesday (July 14). The body would test dangerous capabilities, cybersecurity risks and whether model safeguards can be bypassed
Artificial general intelligence capable of matching the human brain across a broad range of tasks is “probably only a few short years away,” leaving society a “precious window” to establish oversight, Hassabis said, per the report
Hassabis proposed modeling the organization on the Financial Industry Regulatory Authority (FINRA), the self-regulatory organization overseeing U.S. securities firms, the report said
Hassabis said a U.S.-initiated system could become the foundation for international standards as governments confront risks from increasingly capable systems, according to the report
The framework could give AI companies a common testing process while avoiding a separate licensing regime for every model. Independent testers could compare models against common thresholds and update tests as new capabilities emerge
The White House has already moved toward pre-release oversight. A June executive order sought voluntary access to covered frontier models for up to 30 days before release to test advanced cyber capabilities. Google DeepMind, Microsoft and xAI previously agreed to provide models for federal national security testing
The debate intensified after the U.S. restricted access to Anthropic’s Mythos models over concerns that safety controls could be bypassed. Anthropic complied while arguing that the identified jailbreak did not justify a broad recall
Congress is considering additional requirements. A House bill introduced in June would require frontier developers to report dangerous capabilities, breaches and safety incidents within seven days of discovery
Hassabis’ proposal reflects a shift from debating whether frontier AI needs oversight to deciding who conducts the tests, which models qualify and whether regulators can delay deployment
For all PYMNTS AI coverage, subscribe to the dailyAI Newsletter
