Paris Call for AI Safety and Security: What Frontier Model Commitments Mean for Enterprise AI Deployers

On November 9, 2026, French President Emmanuel Macron hosted a meeting on AI and security at the Élysée Palace, and the participants produced a voluntary declaration on frontier AI. Its English title is the Paris Call for AI Safety and Security. The French version, the “Appel de Paris pour la sécurité de l’intelligence artificielle”, has been translated in some reports as the Paris Call for the Security of AI. The text was signed by governments, ten AI companies including Anthropic, Microsoft, Mistral AI and OpenAI, and civil society groups. It commits them to common evaluation standards for the most capable models, faster sharing of serious incidents and clearer accountability. It is not law. For enterprises that buy and deploy frontier models, though, it sets out the safety evidence large developers say they will provide.

What was signed, and by whom

The Élysée’s published text lists the signatories. The governments are Austria, Bahrain, Canada, Croatia, Cyprus, Estonia, Finland, France, Germany, Ireland, Kenya, Latvia, Liechtenstein, Luxembourg, the Netherlands, Norway and Portugal. The companies are Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia. The civil society signatories are LawZero, SaferAI, AI Safety Connect, Ifri, the Geopolitical Center for the Datasphere, GPAI Policy Lab and CeSIA. The page says the statement “remains open to new endorsements and is being updated,” and early press reports gave smaller government counts, so the list may still change.

Attendance and signature were not the same. The Élysée’s French-language communiqué says the meeting included leaders or representatives of Japan, the United Kingdom, Sierra Leone and the US Treasury Secretary, as well as the UN, the OECD and researchers such as Yoshua Bengio. None of those countries appeared on the signatory list when we checked. Google and Meta were also absent from the company list. According to the communiqué, this was the first time AI companies had signed such a text alongside governments.

The declaration builds on earlier work, including the Hiroshima process, the Frontier AI Safety Commitments from the Seoul summit, the G7 meeting in Evian in June and the “Call for Control of Frontier AI Models” that Norway and Finland launched in September and France joined on September 23.

The three areas of consensus

The text records three points of agreement: “the need for a clear, understandable, and predictable regulatory framework; the urgency of achieving tangible and immediate results in transparency, incident reporting, risk management and verification; and the absolute necessity of implementing rigorous evaluation standards.” Some reports quote the last phrase as “assessment standards,” which is a translation variant. The French communiqué put it more briefly: independent evaluators, aligned standards for evaluation and for reporting major incidents, and clearer responsibility for each actor.

Those ideas become three commitments.

• Common evaluation and risk management. Policymakers, researchers and developers will build a shared framework for evaluating frontier models, focused on dangerous misuse in cyber and CBRN domains and on risks from growing autonomy, misalignment and loss of control. The International Network for Advanced AI Measurement, Evaluation and Science (NAAIMES) is named as a vehicle. Non-sensitive evaluation findings, such as red-teaming protocols, benchmarks and stress-test results, would be published in a standard format in a public repository.
• Responsibility and accountability. Each actor is responsible according to its role. Governments say legal frameworks should clarify who is liable for large-scale harm, including across borders. Developers commit to disclose capabilities, limitations and known risks, support qualified independent evaluators where proportionate, maintain fail-safe mechanisms to stop models “including during training and internal use,” address the removal of safeguards from open-weight models, and run internal governance with red-teaming and independent oversight.
• Information sharing and early warning. Signatories will strengthen reporting arrangements, align incident classification, and share timely information on serious cyber and AI control incidents and “material precursor signals” between companies and governments. They will also explore a dedicated cooperation mechanism. Governments commit to secure, confidential whistleblower channels without fear of retaliation.

What the declaration does not do

The Paris Call creates no binding obligations, no regulator and no penalties. Its developer commitments are hedged with phrases such as “where appropriate and proportionate to the risk” and “consistent with existing practices and arrangements.” CeSIA signed it but said it sets no common risk thresholds, guarantees no model access for evaluators and gives public authorities no power to suspend development. The group called for those gaps to be closed. Competition between labs and countries also limits how far voluntary safety work can go, a tension covered in our piece on why AI labs struggle to slow risky model testing.

The Paris Call sits on top of existing law. In the EU, the AI Act’s obligations for general-purpose AI models apply, including incident reporting duties for models with systemic risk. Other countries are building their own institutions, as Australia’s plan for an AI safety regulator with frontier AI obligations shows.

Why enterprise deployers should pay attention

Most companies do not train frontier models. They license them through APIs, cloud platforms and software that embeds them. Their exposure comes from the model supply chain: what the model can do, how it fails and how quickly the provider will tell them when something goes wrong. The Paris Call matters to them because it states publicly what the signing developers will disclose and how they will handle incidents. Buyers can use those statements in vendor reviews and contract talks. That fits the wider privacy risk between a business and its AI models, where the controls a company can rely on are often set upstream.

Questions worth adding to AI vendor due diligence now:

1. Capability and limitation disclosures. Ask for current model or system cards, known risks and intended-use restrictions, and check that your use case fits them.
2. Evaluation evidence. Request red-team and third-party evaluation summaries for the model version you use. Ask whether results will appear in a shared repository like the one the declaration describes.
3. Incident notification. Agree on what counts as a serious incident, how quickly you will be told, through which channel and with what detail. Map that to your own breach and AI incident playbooks so classifications line up.
4. Fail-safe and rollback. Understand what happens to your service if the provider pauses, restricts or withdraws a model. This is the commercial side of the AI kill switch debate, and an exit or fallback plan is a continuity issue as much as a safety one.
5. Open-weight safeguards. If you fine-tune or self-host open-weight models, document which safeguards you keep and which you remove, since the declaration flags safeguard removal as a risk.
6. Whistleblowing and escalation. Make sure staff can raise AI risk concerns internally and externally. Our coverage of non-disparagement clauses and AI whistleblower law shows why contract language matters here.

Governance teams should also log the Paris Call in their regulatory horizon tracking. Voluntary texts like this often feed into later standards and procurement rules, and they can shape what regulators treat as reasonable practice.

Frequently asked questions

Is the Paris Call legally binding?

No. It is a political declaration. Signatory governments may act on it through national law, standards or procurement, and companies may reflect it in their safety frameworks, but the text has no enforcement mechanism.

Does it apply to companies that only use AI models?

Not directly. Its commitments are aimed at frontier developers and governments. Deployers benefit from the disclosures and incident sharing it promises and can ask vendors to honour them in contracts.

Why do the reported signatory numbers differ?

The declaration is open to new endorsements, and the Élysée page has been updated since signing. Early reports listed fewer governments than the 17 shown when we checked.

How does this relate to the children’s AI charter announced in Paris?

They are separate. The Paris Call addresses frontier model safety, security and incident reporting, while the children’s charter deals with protecting young people from AI harms.

A promise you can put in a contract

Declarations do not enforce themselves, and the Paris Call leaves the hard questions on thresholds and access open. Its value for most organisations is more practical. Leading developers have stated publicly what they will disclose, evaluate and report, and buyers can now ask them to put those commitments in writing.