AI companies pledge shutdown controls in Paris declaration without binding enforcement

OpenAI, Anthropic, Microsoft and other AI developers signed a Paris declaration on November 9, 2026, committing to controls that can interrupt models under their control. But the Paris Call for AI Safety and Security creates no binding international enforcement system—and one civil-society signatory is already calling for public powers to suspend development or deployment.

Governments including France, Canada, Finland, Norway and the Netherlands joined companies and civil-society organizations in signing. The political commitment sets neither deadlines nor sanctions. Its published supporter list remains open to endorsements and is being updated.

CeSIA, the Center for AI Security, signed the declaration and welcomed it as a first step. In its November 10 response, it identified three missing guarantees: common risk thresholds, assured model access for evaluators and public authorities, and a suspension mechanism controlled by public authorities.

The declaration does envisage government oversight through safety standards established or recognized by national authorities. CeSIA wants a more specific framework: mandatory independent evaluations and common risk thresholds that determine whether development and deployment may continue.

These decisions must rest with public authorities, who must be able to rely on independent evaluations with guaranteed access to models.

Charbel-Raphaël Ségerie, CeSIA executive director, on suspending development or abandoning deployment

Shutdown controls, within developers’ reach

The shutdown commitment covers model operations within a developer’s control, including training and internal use. The declaration gives restrictions, pauses and shutdowns as examples of fail-safe mechanisms. It is a promise to interrupt operations developers control, not a universal switch for every use of their models.

Developers also commit to disclosing capabilities, limitations and known risks. Support for qualified independent evaluation is conditional: it should be appropriate, proportionate to risk and consistent with existing practices and arrangements.

Inside companies, the text calls for continuous risk management, real-time monitoring of emerging capabilities and rigorous adversarial testing, often called red-teaming. Developers commit to working toward governance with defined responsibilities, adequate resources and independent oversight, alongside reporting lines that protect internal researchers.

Comparable tests and shared warnings

A proposed common evaluation framework would use shared standards, tools and methods for frontier models—the most advanced AI systems. Its focus includes dangerous cyber misuse and risks from growing autonomy or models behaving contrary to intended goals, including loss of control.

Signatories agree to publish non-sensitive evaluation findings in a standardized format and create a publicly accessible shared repository. Potential contents include adversarial-testing protocols, benchmarking tools, risk assessments and stress-test results.

Incident cooperation would extend beyond completed harm to significant warning signals. The text calls for timely sharing of serious cyber and AI-control incidents and work toward common incident classifications. Public officials also commit to secure, confidential channels for whistleblowers to raise reasonably grounded systemic-risk concerns without fear of legal or professional retaliation.