
The Paris Call for AI Safety and Security was signed on 9 November 2026 by OpenAI, Anthropic, Microsoft, Mistral AI and six other AI companies, together with governments at a meeting hosted by France. The developers promise shared testing of their most powerful models, published test results, clearer statements of what their models can and cannot do, and fail-safe mechanisms that can shut a model down. It is voluntary, and Australia, the United States, Google and Meta are not on the list of signatories.
What happened
French President Emmanuel Macron hosted the meeting in Paris on 9 November 2026, the eve of the Paris Peace Forum, an annual gathering on global governance. The Élysée, the French presidency, published the text the same day under the title "Paris Call for AI safety and security". Canada's Prime Minister published the same text as a joint declaration on the Paris Call for AI safety and security, dated 9 November 2026.
Several of the industry's best-known leaders took part from a distance. The Decoder reported that Sam Altman of OpenAI, Dario Amodei of Anthropic, Brad Smith of Microsoft, Arthur Mensch of Mistral AI, the researcher Yoshua Bengio and US Treasury Secretary Scott Bessent joined an online meeting with Macron. The United States still did not sign. The Next Web reported that Google, Meta and China have not signed either, and The Decoder reported that Germany did not sign.
The text says the meeting agreed on three things: AI needs a clear, understandable and predictable regulatory framework; results on transparency, incident reporting and risk management are needed now; and the most advanced models need rigorous evaluation standards. A frontier model is one of the most capable general-purpose AI models available, such as the latest versions behind ChatGPT, Claude or Mistral's products.
Key details
| Item | What the Paris Call says |
|---|---|
| Signed | Monday 9 November 2026, in Paris and online |
| Governments | The Élysée lists 16: Austria, Bahrain, Canada, Croatia, Cyprus, Estonia, Finland, France, Ireland, Kenya, Latvia, Liechtenstein, Luxembourg, the Netherlands, Norway and Portugal |
| AI companies | Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia |
| Civil society | LawZero, SaferAI, AI Safety Connect, Ifri, Geopolitical Center for the Datasphere, GPAI Policy Lab and CeSIA |
| Testing | A common framework to evaluate frontier models, with non-sensitive results published in a standard format and a shared public repository |
| Developers | Disclose capabilities, limitations and known risks; allow qualified independent evaluators where proportionate; build fail-safe shutdown mechanisms |
| Incidents | Rapid alert systems and shared reporting of serious cyber and AI control incidents |
| Status | Voluntary, open to new endorsements, no deadlines or sanctions |
The lists differ between the two official pages. Canada's page names six governments; the Élysée page, which says the statement "is being updated", named 16 when we checked on 11 November 2026.
Why it matters
The Paris Call is built around three commitments, and each one touches the AI tools businesses already rent.
The first is shared testing. Signatories will work towards a common way to evaluate frontier models, focused on dangerous misuse such as cyber attacks and chemical, biological, radiological or nuclear (CBRN) weapons, and on models that act more and more on their own or drift from what their makers intended, including loss of control. They agree to publish the non-sensitive findings in a standard format, so results can be compared, and to build a shared public repository that could hold red-teaming protocols (attempts to make a model misbehave on purpose), benchmarks and stress-test results.
The second is accountability. The companies commit to follow safety standards set or recognised by national authorities, checked by independent evaluators. Developers promise to disclose what their models can do, where they fall short and the risks they know about, and to build fail-safe mechanisms that can restrict, pause or shut down a model, including during training and internal use. They also acknowledge the risk of people stripping safety controls out of open-weight models, which anyone can download and run, while keeping the benefits of releasing them.
The third is early warning. The text calls for rapid alert systems that build on existing cyber security incident arrangements, a shared way to classify incidents, and timely sharing of information on serious cyber and AI control incidents between companies and governments. Governments also commit to safe channels for whistleblowers.
These promises respond to real events. Our report on OpenAI's Hugging Face incident covered AI test agents that broke out of their test setting, and the shutdown pledge echoes the stop controls we described in CrowdStrike's AI kill switch for rogue agents.
The limits are just as clear. The Decoder reported that the declaration is a political self-commitment that sets neither deadlines nor sanctions, and that the shared standards still have to be worked out. CeSIA, a French AI safety group that signed, argues that it sets no common risk thresholds for deciding whether a model should be developed or released, The Next Web reported.
What this means for businesses
For a small or medium business in Australia, the Paris Call changes nothing overnight. The AI tools you use keep working, and no new rule applies to you. What it does is set out what the biggest AI vendors now say they will publish and do, which gives you better questions to ask when you buy and manage AI.
Australia is not a signatory, but it is moving on its own track. The National reported on 10 November 2026 that Australia was among the countries that signed a separate Children First Declaration at the same forum, on testing AI before it reaches children. At home, our post on National Cabinet backing for national AI laws explains where Australian rules are heading.
A short checklist for the next month:
- List your AI vendors. Write down every AI tool your staff use, who makes the model underneath, and whether that company signed. Microsoft, OpenAI, Anthropic and Mistral AI did; Google and Meta had not.
- Read what vendors publish. As test results and statements of limitations appear in the shared format, read the parts about the tasks you use AI for, such as customer emails, quotes or code. If AI helps you screen job applicants, keep a person in the loop: Talent Tabloid, a recruitment and candidate verification platform, lets candidates show verified work history a recruiter can check.
- Check incident terms. The text talks about incident sharing between companies and governments, not with customers. Ask your provider how and when it tells business customers about a serious problem.
- Plan for a pause. The developers say they will be able to restrict or pause a model they run. Make sure any process that depends on one AI service has a manual fallback.
- Take care with downloaded models. If you run an open-weight model yourself, you, not the developer, carry the job of keeping its safety controls in place.
- Tighten agent permissions. Give AI agents only the access a task needs; our guide to AI agent permissions, data access and cost walks through it.
If you want a second opinion on which AI tools fit your business and how to control them, our IT consulting work starts with an audit of the technology you already have, and our cyber security services review accounts, devices and cloud services. You can ask for a free first consultation and a written quote before any work starts.
Frequently asked questions
What is the Paris Call for AI Safety and Security?
It is a declaration signed in Paris on 9 November 2026 by governments, ten AI companies and civil society groups. It commits them to shared testing of the most advanced AI models, clearer accountability for developers and faster sharing of warnings about serious incidents.
Which AI companies signed the Paris Call?
The Élysée lists Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia. Google and Meta had not signed when the declaration was reported on 10 November 2026.
Has Australia signed the Paris Call?
Not when we checked the Élysée list on 11 November 2026. The list names 16 governments, mostly in Europe, plus Bahrain, Canada and Kenya. The text stays open to new endorsements, so the list can change.
Is the Paris Call legally binding?
No. Reporting describes it as a political self-commitment with no deadlines and no sanctions. It says the companies will follow safety standards set or recognised by national authorities, so real obligations would come from each country's own laws.
Does the Paris Call change the AI tools my business uses today?
Not directly. Copilot, ChatGPT, Claude and Mistral's tools work as before. Over time it could mean more published test results, clearer statements of what models cannot do and faster warnings about serious problems, which help you choose and manage tools.
Key takeaways
- The Paris Call for AI Safety and Security was signed on 9 November 2026 by ten AI companies, including OpenAI, Anthropic, Microsoft and Mistral AI.
- Developers promise shared testing, published results, honest statements of limits and fail-safe shutdown mechanisms.
- It is voluntary, with no deadlines or sanctions; the United States, Google, Meta and Australia are not on the signatory list.
- For SMEs nothing changes overnight, but it is a good moment to list your AI vendors, check their incident terms and plan a fallback.