Google DeepMind CEO Demis Hassabis is calling on the US-led standards body to independently test the world’s most advanced artificial intelligence models against national security threats before they are released.
The proposal would cover frontier models developed in the United States and abroad, including open and closed systems, the Financial Times reported on Tuesday (July 14). The agency plans to test whether dangerous features, cybersecurity risks, and safety measures in the model can be circumvented.
The report says artificial general intelligence on par with the human brain across a wide range of tasks is “probably within just a few years,” leaving a “precious window” to establish surveillance in society, Hassabis said.
Hassabis proposed modeling the organization on the Financial Industry Regulatory Authority (FINRA), the self-regulatory body that oversees U.S. securities firms, the report said.
Hassabis said the U.S.-led system could become the basis for international standards as governments face risks from increasingly sophisticated systems, according to the report.
The framework could provide AI companies with a common testing process while avoiding the need for separate licensing regimes for each model. Independent testers can compare models to common thresholds and update tests as new features emerge.
The White House is already moving toward pre-release monitoring. A June executive order required voluntary access to eligible Frontier models for up to 30 days prior to release to test advanced cyber features. Google DeepMind, Microsoft, and xAI previously agreed to provide models for federal national security testing.
The debate intensified after the United States restricted access to Anthropic’s Mythos model over concerns that security controls could be circumvented. Anthropic responded even though it maintained that the identified jailbreaks did not warrant a major recall.
Congress is considering additional requirements. A House bill introduced in June would require frontier developers to report unsafe features, violations and safety incidents within seven days of discovery.
Hassabis’ proposal reflects a shift from discussing whether frontier AI requires oversight to deciding who will conduct the tests, which models are eligible, and whether regulators can delay implementation.
For all of our coverage of PYMNTS AI, subscribe to our daily subscription AI Newsletter.
