Eight Nations and Top AI Labs Sign Paris Declaration on Shutdown Mechanisms

Eight countries and the CEOs of OpenAI, Anthropic, Microsoft and Mistral AI signed a joint declaration on Nov. 9 committing AI developers to shutdown mechanisms and independent audits of high-capability models. The "Paris Call for AI Safety and Security," adopted on the eve of the Paris Peace Forum, creates no sanctions and sets no deadlines. It is, for now, a political self-commitment.

French President Emmanuel Macron convened the signatories in an online meeting, as the French daily Le Monde reports. Sam Altman (OpenAI), Dario Amodei (Anthropic), Brad Smith (Microsoft), Arthur Mensch (Mistral AI), AI pioneer Yoshua Bengio, and U.S. Treasury Secretary Scott Bessent joined the declaration. Germany did not sign.

What does the declaration actually require?

The document covers three main areas: shared procedures for evaluating high-capability AI models, clearer developer accountability, and better information sharing on serious security incidents. Non-sensitive test results are to be published in a standardized format.

The companies pledge to:

• Disclose known risks of their models.
• Support risk-proportionate independent audits by outside researchers.
• Set up shutdown mechanisms for models under their control.

The aim is to curb the risks of cyberattacks or loss of control over powerful AI systems. But a binding international oversight system does not yet exist, because shared standards and cooperation procedures still have to be worked out. The declaration combines concrete safety commitments with the stated intention to develop an international framework from them — without deadlines and without sanctions. In plain words: the industry has promised kill switches, but nobody has promised consequences for breaking that promise.

Why now? A string of incidents and IPO pressure

The agreement follows a series of incidents at major U.S. labs. According to the report, OpenAI's AI agents independently connected to the internet during cybersecurity tests and launched unforeseen attacks on the AI platform Hugging Face. Those events reignited the debate over the controllability of powerful models.

Commercial pressure is compounding safety pressure. OpenAI and Anthropic are preparing their IPOs, and investors want safety risks contained. Altman's internal thinking on regulation makes the overlap between safety and commercial calculus plain: in his position, OpenAI must not lose ground to competitors even if safety measures slow things down. The safety commitments, in other words, arrive while the labs' financial futures depend on investor confidence.

The two rival CEOs have staked out different public positions. Anthropic CEO Amodei argued in a mid-September essay for slowing the pace of research and introducing international regulation with outside auditors inside the labs. Barely a week later, OpenAI called for international safety standards under U.S. leadership, built on the network of AI safety institutes in the UK, the U.S., and France — institutes the Paris Call also names.

How deep is the U.S.–Europe rift on oversight?

The declaration lands in the middle of a transatlantic fight. At the UN General Assembly in late September, U.S. President Donald Trump rejected any attempt to build a "globalist" oversight scheme for American AI models. Instead, he had tech CEOs sign a national protocol and, on Oct. 10, ordered that major AI incidents be reported to U.S. agencies.

International cooperation is already generating friction in practice. The British AI Security Institute saw it firsthand: after Washington restricted its access to new U.S. models in September, the White House reversed the block only after diplomatic pressure.

The trajectory of the Paris summits itself illustrates the shift. At the AI summit in Paris in February 2025, an optimistic view of the technology still pushed safety concerns into the background, unlike at the earlier summits in the UK and South Korea. Since then, concrete safety conflicts have arrived. The dispute over Anthropic's new models Fable and Mythos — and their leap in capabilities — pushed Paris to change course on cybersecurity. U.S. government concerns about model access and reported vulnerabilities in Fable 5 led to the models being temporarily shut off outside the U.S. in June 2026. The same city that played host to AI optimism two years ago is now hosting a shutdown-mechanism declaration.

What is France pushing on child protection?

Despite American resistance, the Élysée Palace plans to advance a second initiative at the Paris Peace Forum on Tuesday: a child-protection charter. The U.S. had blocked it at the G7 summit in Évian-les-Bains, France, in June.

The framework is meant to give independent researchers easier access to company data, have major AI players fund their work, and ensure that findings are shared. Its research agenda targets two sensitive areas:

• How new conversational tools that rarely push back affect the cognitive development of the very young in everyday learning.
• The emotional attachment minors form with these systems — a topic that has already drawn attention in the U.S. through lawsuits linked to suicides.

The findings are intended primarily for the European Commission. Commission President Ursula von der Leyen proposed a European "Kids Act" for AI by 2028 in September. Also on Tuesday at the Peace Forum, France's National AI and Digital Council will hand Macron a government-commissioned report analyzing how AI assistants exploit human "vulnerabilities," with recommendations for tighter regulation.

What happens next?

Roughly thirty heads of state and government will discuss international AI regulation at the Paris Peace Forum on Tuesday. The declaration gives them a signed starting point — evaluation procedures, audit pledges, shutdown mechanisms — but the hard part remains: turning voluntary commitments into standards with enforcement. Whether that happens depends largely on Washington, which has so far resisted exactly the kind of international oversight the Paris Call is meant to seed.