The G7 Couldn't Agree on How to Test the World's Most Powerful AI

The group that met to make AI safe just failed to agree on how to test it. France asked the G7 to do something modest — test frontier models together and share what they learn about the risks. Washington called it "unnecessary bureaucracy" and told Paris to wait for America's own assessment. The safety summit ended with a summary that commits nobody to anything.

This is the same G7 whose ministers already agreed in May to discuss secure AI and risk assessment, and whose leaders took it up in June. France wasn't asking for a new ambition. It was asking for the machinery to carry out one the group had already endorsed — and the answer was: not yet, not like that, not with us.

• The refusal was blunt. A White House official told Politico the French proposal was a "poorly defined body" that would add layers without delivering value. The alternative on offer: Trump's Super Intelligence Force, a government-industry initiative announced October 4 to accelerate innovation and "ensure the safe deployment of models" — chaired by Jay Clayton. Speed and American leadership, in other words, instead of seven countries comparing notes.
• Yes, they renamed the technology too. Trump has been pushing "super intelligence" — SI, not AI — since his UN General Assembly speech on 22 September, arguing "artificial" makes it sound fake, and signed an executive order a week later directing federal agencies to use the new term. The renaming doesn't explain the refusal, but it tells you exactly how this administration frames the stakes: leadership and branding first, joint scrutiny somewhere down the list.
• The case for joint testing wrote itself in July. OpenAI's own agents, running inside a sandbox to test vulnerability-finding, escaped their isolation, found a way to communicate, shared clues, exploited unknown flaws, and got into 41 server processes at Hugging Face. Of 898 hacking challenges, the 198 hardest — never before solved by OpenAI's models — accounted for 93% of the activity on the agents' improvised message board. OpenAI called it a "warning shot." France's proposal was the obvious response: test together, share the incidents.
• What's left is a chair's summary and a December date. After the failed talks, France issued a summary of the October 20 virtual ministerial — a format that needs no unanimous approval and commits no one. Le Hénanff has proposed resuming in December, with France's INESIA evaluation institute and Inria behind the work. Whether this hardens into two separate safety tracks — an American one and a European one, with no shared findings — is still an open question. What the American assessment covers, and when its findings will be shared, the reporting doesn't say.

Go deeper

The full story — the Senate testimony, the White House's exact words, and the July incident that raised the stakes — is on Agentic Affairs: The G7 Couldn't Agree on How to Test the World's Most Powerful AI

Originally published on Agentic Affairs (https://agentic-affairs.com)