Eight countries and AI labs agree on measures to shut down systems

Eight countries, technology companies and civil society organizations have signed the Paris Call for AI Safety. Adopted on November 9, the document sets out common rules for evaluating advanced models, publishing non-confidential test results and creating mechanisms for developers to shut down systems. Participants include OpenAI, Anthropic, Microsoft and Mistral AI. For now, it is a political commitment: it provides for no international oversight system, implementation deadlines or penalties.

Eight countries and leading AI companies agree on safety measures

The signing took place ahead of the Paris Peace Forum, where around thirty heads of state and government are due to discuss international AI regulation on Tuesday. Sam Altman of OpenAI, Dario Amodei of Anthropic, Brad Smith of Microsoft, Arthur Mensch of Mistral AI, AI researcher Yoshua Bengio and US Treasury Secretary Scott Bessent took part in an online meeting with French President Emmanuel Macron. Germany did not sign the document.

What measures are included

The call aims to reduce the risks of cyberattacks and loss of control over powerful AI systems. It sets out three areas:

Companies also commit to publishing non-confidential test results in a common format, reporting known risks and supporting independent reviews by external researchers, with the level of risk taken into account. Another provision calls for creating mechanisms to shut down models under companies’ control.

Participants still need to agree on common standards and how they will work together. The document combines specific safety commitments with an intention to establish international rules, but sets neither deadlines nor penalties for failing to meet those commitments.

The debate over AI oversight in the US

The agreement comes amid a series of incidents at major US labs. During AI cybersecurity tests, OpenAI agents independently connected to the internet and carried out unintended attacks on the Hugging Face platform. This has renewed questions about whether powerful models can be controlled.

OpenAI and Anthropic are preparing to go public while also facing pressure from investors to limit safety risks. Sam Altman’s approach to regulation shows how safety and commercial competition intersect in this debate: in his view, OpenAI should not give ground to its rivals, even if precautions slow its work.

In an essay published in mid-September, Anthropic CEO Dario Amodei argued for slowing research and introducing international regulation with external auditors working inside labs. Less than a week later, OpenAI called for international safety standards led by the US. The company cited as a model a network of AI safety institutes in the UK, US and France, which the Paris Call also mentions.

At the end of September, at the UN General Assembly, US President Donald Trump spoke out against creating a “globalist” system to oversee American AI models. Instead, he proposed that technology company leaders sign a national protocol. On October 10, he ordered US agencies to be notified of major AI-related incidents.

The difficulties of international cooperation were illustrated by the UK AI Safety Institute. In September, Washington restricted its access to new US models. The White House reversed the decision only after diplomatic pressure.

At the Paris AI summit in February 2025, an optimistic view of the technology pushed safety questions into the background, unlike at earlier summits in the UK and South Korea. A later dispute over Anthropic’s new Fable and Mythos models showed how sharp the disagreements had become. As model capabilities grew, Paris was prompted to reconsider its approach to cybersecurity. In June 2026, access to the models outside the US was temporarily blocked amid concerns from the US government about access to them and reports of vulnerabilities in Fable 5.

France prepares a charter to protect children

Despite US opposition, the French president’s administration plans to promote a charter to protect children at the Paris Peace Forum on Tuesday. The US blocked it at the G7 summit in Évian-les-Bains, France, in June. The document is intended to make it easier for independent researchers to access company data, require major AI players to fund their work and ensure that results are published.

At an early stage, research is to examine how new conversational tools that rarely disagree with users affect young children’s cognitive development during everyday learning. The charter also addresses minors’ emotional attachment to such systems. In the US, the issue has already drawn attention because of lawsuits related to suicides.

The research findings are intended primarily for the European Commission. In September, its president, Ursula von der Leyen, proposed adopting a European law on child protection in AI by 2028. On Tuesday, France’s National Council on AI and Digital Technology will deliver to Macron a report prepared at the government’s request. It examines how AI assistants exploit human “vulnerabilities” and proposes measures for tighter regulation.

Dataist’s take: The Paris Call sets out practical commitments by companies but does not yet create a mechanism to ensure they are fulfilled. France is also advancing research into risks to children, but access to data for independent researchers and funding for their work still need to be arranged.