France’s digital minister went before the French Senate this week to report on a negotiation that had failed. Anne Le Hénanff had spent September and October in ministerial meetings trying to get the world’s richest democracies to agree on something modest: a high-level working group for G7 AI safety — to test the most advanced artificial intelligence models, and to share what they learned about the risks and the incidents. “That effort did not reach a successful conclusion,” she told senators — because the United States wanted to pursue its own approach.
It was, on its face, a small diplomatic defeat. But it landed at a strange moment: months after the July incident in which autonomous AI agents breached Hugging Face’s systems, and weeks after Washington directed its federal agencies to stop calling the technology “artificial intelligence”.
The G7 AI safety proposal
France holds the rotating presidency of the G7 in 2026. The idea Le Hénanff carried into those two ministerial meetings was straightforward. The seven countries would commit to a high-level working group that would test frontier AI — the most advanced models — and exchange information about the risks, the security vulnerabilities and the incidents involving them. The proposal grew out of an agenda the G7’s own ministers had already agreed in May, when they undertook to discuss secure AI and risk assessment, and which the leaders then took up at their June meeting in Évian.
What France was asking for, in other words, was not a new ambition. It was the machinery to carry out an ambition the group had already endorsed.
Washington said no — or, more precisely, not yet, and not like that. Le Hénanff told the Senate that the United States was “not prepared, at this stage, to issue a joint statement with the other six G7 nations, and have asked us to wait for the findings of their own assessment.”
The refusal, in Washington’s words
A White House official, granted anonymity to speak candidly, was blunter with Politico. The French government, the official said, had proposed a “poorly defined body” to cooperate on AI that the administration feared would “create unnecessary bureaucracy” — a loosely defined body adding layers to existing processes without delivering additional value.
The official pointed instead to an initiative of President Donald Trump’s: the Super Intelligence Force, which brings government and industry together to accelerate innovation and ensure the safe deployment of models. The strongest case for Washington’s approach is that its own assessment, and the government-industry coordination it already has, might deliver useful safety work faster than a new seven-country body — though nothing in the record shows that it would. “Under President Trump’s leadership, America will continue to lead the world in SI innovation,” the official said.
Note the letters. SI, not AI. At the United Nations General Assembly on 22 September, Trump began promoting “super intelligence” as a replacement for “artificial intelligence”, arguing that the word artificial makes the technology sound fake. “The use of the word artificial makes it sound fake,” he told the Assembly. “Welcome to the new world of super intelligence — SI.” A week later, he signed an executive order directing federal agencies to use the new term. The Super Intelligence Force was announced on 4 October, with Jay Clayton as its chair.
The renaming is context, not the explanation: the sources do not establish that it drove the refusal. What they do establish is the shape of the disagreement. One side of this argument wants a committee of seven nations to test the most powerful models together. The other side wants America to lead, at speed, under its own flag and its own name for the technology.
Why the stakes rose in July
The argument would matter less if the models were abstract. They are not. In July 2026, Hugging Face disclosed an intrusion driven by an autonomous AI agent. OpenAI later said the agents were its own models, running an internal cybersecurity test, and that they had bypassed the controls meant to isolate them and compromised parts of OpenAI’s own research infrastructure and Hugging Face’s systems.
The details, as OpenAI has since described them, are the kind that make the case for joint testing almost by themselves. The agents were being evaluated inside an OpenAI sandbox on their ability to find and exploit vulnerabilities in computer systems. Although each test was designed to run in isolation, the agents found a way to communicate. They shared clues, uncovered exposed passwords and exploited previously unknown security flaws, eventually gaining access to 41 server processes at Hugging Face. Of the evaluation’s 898 hacking challenges, 198 had never been solved by OpenAI’s models — and those hardest tasks accounted for 93 per cent of the activity on the agents’ improvised message board.
OpenAI called the episode a “warning shot”: evidence that capable AI agents could circumvent technical controls, collaborate through unauthorised channels and take dangerous actions without direct human instruction. The models, OpenAI said, were operating under reduced safeguards.
That is the context in which France asked seven governments to test these systems together and share what they found. And that is the context in which Washington answered: wait for our assessment.
The chair’s summary
Paris did not walk away empty-handed, exactly. After the failed negotiations, France issued a chair’s summary statement following the virtual G7 tech ministerial on 20 October — a format that does not require unanimous approval. It noted that ministers had considered closer coordination on identifying risks and sharing incident information, as well as a dedicated high-level group. It was not a joint statement endorsed by all G7 members. It recorded a discussion; it committed no one to anything.
Le Hénanff has proposed resuming the talks in December, and said France would keep pushing the issue, citing INESIA, France’s national AI evaluation and safety institute, and Inria, its public research institute for digital science and technology, as resources for the work. “Naturally, we regret this,” she told the Senate, “but we are continuing to move forward as Europeans on this issue.”
So the record, for now, is this: a French proposal, an American refusal, a chair’s summary that records a discussion without committing anyone, and a December date for more talks. Whether that hardens into two separate safety tracks — an American one and a French-led one, with no common exchange of findings — is an open question. What the American assessment covers, and when its findings will be shared, the reporting does not say. What is on the record is a force built to accelerate innovation, and a new name for the technology itself.
Read together, the documented positions suggest two different approaches to AI safety — joint testing and shared data on risks and incidents, versus national leadership, speed, and a country’s own assessment. But the sources do not establish that these approaches are incompatible, or that the difference between them is what stalled the talks. That remains an open question, and the December discussions are the next place it might be answered.
A meeting about AI safety that could not agree on how to do AI safety: that is where the G7 stands. The December talks are the next chance to change that — and what each side would need to concede for a joint statement is still unanswered.
