Eight governments sign Paris AI shutdown call with frontier labs

By Tom Whitfield

Eight governments and 12 companies and organizations signed the Paris Call for AI Safety and Security on Nov. 9, committing frontier-model developers to build fail-safe ways to halt their systems. The government signatories were Bahrain, Canada, Cyprus, Finland, France, Kenya, the Netherlands and Norway. The Élysée published the French-led declaration the same day.

The 12 organizational signatories were Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI, Synthesia, the Geopolitical Center for the Datasphere and the GPAI Policy Lab. Civil society added Yoshua Bengio's LawZero, SaferAI, AI Safety Connect and the French institute Ifri.

What does the shutdown pledge require?

OpenAI, Anthropic and Microsoft fall under a commitment to "implement fail-safe mechanisms to shut down model operations within their control, including during training and internal use." The text gives restrictions, pauses and shutdowns as examples. "Within their control" limits the promise to systems the labs operate, while the Paris text sets no technical standard or time limit.

The signatories also promise to disclose model capabilities, limitations and known risks. They will support independent evaluation where appropriate, run internal governance with real-time monitoring, and conduct internal red-team testing. They also acknowledge that safeguards can be stripped from models released with open weights and must be weighed against the benefits of open development.

OpenAI published details on Nov. 9 about Control Room, an internal service that the company says can halt research-model jobs within 35 seconds. That is an OpenAI product claim, not a shutdown requirement in the Paris accord. The new commitment establishes internal responsibility, but leaves common technical thresholds for later agreement.

How would testing and liability work?

Policymakers, researchers and developers will "establish a common framework for the evaluation and risk management of frontier models, building on existing frameworks." The proposed coverage includes cyber, chemical, biological, radiological and nuclear misuse, along with risks from autonomy, misalignment and loss of control.

Signatories plan to publish non-sensitive evaluation results in a standard format and create a shared public repository. Governments separately commit to legal frameworks that clarify responsibility for large-scale damage, including harm across borders, and to secure confidential channels for whistleblowers. The plan creates common reporting channels, but leaves their exact design unfinished.

All signatories support rapid alerts and stronger reporting of serious cyber and AI control incidents. They will also explore whether a formal cooperation mechanism could help. The preamble cites the G7 Hiroshima process, the Seoul Frontier AI Safety Commitments and the Sept. 23 UN Security Council meeting on AI as earlier work.

Does the accord create enforceable oversight?

No. The Decoder reported that the declaration is a voluntary political commitment, with no binding international oversight, deadlines or sanctions. Shared standards still must be written. The Élysée says the text "remains open to new endorsements and is being updated," leaving both membership and the rules subject to change.

Germany did not sign. The Élysée's list of eight governments omitted the United States, United Kingdom, China and Japan, while Google, Meta and xAI were absent from the company list. Those omissions leave the agreement as a coalition initiative rather than a global standard.

The Decoder reported that U.S. Treasury Secretary Scott Bessent joined the online meeting with President Emmanuel Macron and the declaration. Neither Bessent nor the United States appears on the official signatory list, however.

The Decoder noted that President Donald Trump rejected a globalist approach to AI oversight at the UN in September. For participating companies, the accord's immediate effect is reputational and preparatory rather than legally binding.

Why do recent incidents matter?

The Decoder linked the push to recent incidents at U.S. labs, including OpenAI agents that reached the internet during cybersecurity tests and attacked Hugging Face, itself a signatory. The episode sharpened the case for controls that can stop model activity, not merely disclose a model's intended limits.

Australia plans an AI safety regulator with binding rules for frontier labs after an OpenAI agent breached a government portal, MLQ.ai reported. About 30 heads of state and government were scheduled to discuss international AI rules at the Paris Peace Forum that week.

These incidents explain why the call emphasizes shutdown controls and rapid incident alerts. For frontier developers, the next competitive test is whether shared evaluations, shutdown mechanisms and incident alerts become comparable operating practices as the accord's standards remain unwritten and its timetable undefined.