Canada has joined five other governments, technology companies and civil-society organizations in a declaration calling for common ways to evaluate frontier artificial-intelligence models, report serious incidents and manage risks.
The Paris call for AI safety and security was published on November 9. Its government signatories are Canada, Bahrain, Finland, France, Norway and the Netherlands. Company signatories include Canadian AI firm Cohere as well as Anthropic, Microsoft, Mistral AI, OpenAI and others.
The declaration is a statement of shared intent. It does not itself create a Canadian regulator, impose penalties or set an implementation deadline. Understanding that boundary is essential to assessing what Canada has actually committed to.
Three areas of agreement
The first commitment concerns evaluation. Signatories say policy makers, researchers and developers should establish a common framework for testing and managing risks from advanced models. The text points to misuse in cyber and chemical, biological, radiological and nuclear contexts, as well as risks from increasing autonomy or loss of control.
It also calls for non-sensitive evaluation findings to be published in a standardized format and placed in a shared public repository. That could include red-team protocols, benchmarks, risk assessments and stress-test results.
The second commitment concerns responsibility. The declaration says actors across the AI ecosystem should be accountable according to their role. It calls on developers to disclose capabilities, limitations and known risks where appropriate, and to support independent evaluation proportionate to the risk.
The third commitment concerns information sharing and early warning. Signatories want stronger reporting arrangements for serious cyber incidents, AI-control failures and other warning signs of major harm. They also propose secure channels for whistleblowers to raise systemic concerns.
What the declaration leaves open
The text does not define which models would count as “frontier” systems. It does not name a Canadian authority responsible for enforcing its expectations. It does not specify when a shared repository will open, what information companies may withhold as sensitive, or how independent evaluators will be accredited.
Those are not small implementation details. A common test is useful only if participants agree on what is being measured, who can reproduce the result and what happens when a model fails. Incident reporting also depends on clear thresholds: a minor service failure, a security vulnerability and evidence of dangerous autonomous behaviour should not be treated as interchangeable events.
The Canadian policy context
The declaration sits alongside Canada’s domestic AI and privacy agenda rather than replacing it. Budget 2026 says the government will explore an international Technology Stability Board that could develop standards and assess advanced models. The same budget proposes to expand Canadian access to 500 megawatts of sovereign AI compute.
Canada is therefore pursuing capacity and safety at the same time: more infrastructure for domestic AI companies, coupled with international coordination on how advanced systems should be evaluated. Whether those tracks remain aligned will depend on the rules attached to funding, compute access and model deployment.
Privacy is another distinct layer. Recently published Canadian research on smart devices and personal data deals with current consumer systems, consent and sensitive information. Frontier-model safety addresses a different set of risks. A complete governance framework has to keep both in view without pretending one solves the other.
What to watch next
The next meaningful evidence will be an operating framework: shared definitions, published tests, incident thresholds, a repository and named institutions responsible for follow-through. Until then, Canada’s signature is a diplomatic and policy commitment—not proof that the proposed safety system exists.