OpenAI, Anthropic and Microsoft have joined a group of international AI developers and governments in signing a declaration that commits participants to stronger safety controls for advanced artificial intelligence systems, including mechanisms to stop models when necessary.
The Paris Call for AI Safety and Security was signed on November 9, 2026, following discussions convened in Paris by French President Emmanuel Macron. The declaration brings together six governments, 10 technology companies and four civil society organisations, according to the Canadian prime minister’s office, which published the full text.
The government signatories are Bahrain, Canada, Finland, France, Norway and the Netherlands. The companies include OpenAI, Anthropic, Microsoft, Mistral AI, Cohere, Hugging Face, Black Forest Labs, Sakana AI, Synthesia and Domyn.
Among the civil society organisations are LawZero, led by prominent AI researcher Yoshua Bengio, and SaferAI.
The agreement focuses on three areas: evaluating advanced AI models, establishing accountability for their developers and improving the exchange of information about serious incidents.
What the companies have agreed to
Under the declaration, developers of frontier AI models commit to implementing fail-safe mechanisms that can restrict, pause or shut down model operations under their control, making AI Kill Switches a need for the companies signing the agreement. The requirement explicitly covers training and internal use, rather than applying only to publicly released products.
The document states that developers must:
- Disclose their models’ capabilities, limitations and known risks.
- Support independent assessments by qualified evaluators where appropriate.
- Establish mechanisms to stop model operations when safety requires it.
- Monitor emerging capabilities and conduct internal red-team testing.
- Maintain governance structures with defined responsibilities and independent oversight.
The commitments also address the possibility that developers or users could remove or circumvent existing safeguards. The declaration specifically mentions abliteration, a technique used to modify models by removing or weakening particular behavioural restrictions.
On testing, signatories intend to develop common technical standards and evaluation methods. They also agree to publish non-sensitive findings in a standardised format and establish a publicly accessible repository containing material such as benchmarking tools, risk assessments and stress-test results.
The proposed framework will pay particular attention to cyber threats, chemical, biological, radiological and nuclear risks, and the possibility that increasingly autonomous systems could escape human control.
A system for reporting serious incidents
The third commitment concerns information sharing.
Governments, researchers and AI developers have agreed to strengthen existing reporting arrangements and develop faster warning systems for serious cybersecurity incidents and AI control failures.
The declaration calls for better coordination between companies and governments, between governments themselves, and among AI developers.
However, the signatories have stopped short of establishing a fully defined international incident-reporting organisation. Instead, they have agreed to examine whether a dedicated cooperation mechanism is needed and what its responsibilities should be.
Governments have also committed to providing confidential channels through which employees and other individuals can report systemic AI risks without fear of professional or legal retaliation.
However, the initial signatory list is smaller than the group of governments and technology companies that participated in the Paris discussions.
The French presidency said Macron brought together representatives of eight countries, including Japan, the United Kingdom and Sierra Leone, as well as the US Treasury secretary and representatives of the United Nations and OECD.
But the Canadian government’s published list of signatories names only six countries: Bahrain, Canada, Finland, France, Norway and the Netherlands.
The United States and China are not listed as government signatories yet. The UK and Japan participated in the discussions but are also absent from the initial government signatory list, as per our sources.
Some major AI companies like Google and Meta are missing from the published list of company signatories, which includes 10 developers.
No penalties for failing to deliver
The declaration establishes political commitments rather than a new international law.
Its published text does not set legally binding implementation deadlines, financial penalties or sanctions for companies that fail to introduce the promised controls. Nor does it establish an international authority with powers to inspect AI developers and enforce compliance.
Some of the language also leaves important work unfinished. The common evaluation framework still needs to be developed, the shared repository has yet to be established, and the proposed incident-cooperation mechanism remains under discussion.
The signatories have therefore agreed on what they want to build, but the declaration does not establish a completed system for enforcing those commitments.
The French presidency said the declaration remains open to additional endorsements.
France has presented the agreement as part of a continuing international effort to coordinate AI safety standards, following previous initiatives including the 2024 Seoul AI Summit’s Frontier AI Safety Commitments and subsequent G7 discussions.
The Paris declaration adds a specific commitment to shutdown mechanisms and a shared approach to evaluating advanced models.
