Security

Paris Call for AI Safety and Security: Governments and Ten AI Developers Sign Voluntary Pledges on Testing, Shutdown and Incident Sharing

A declaration agreed at the Élysée on 9 November commits signatories to a common framework for evaluating frontier AI risks, fail-safe shutdown mechanisms and faster sharing of serious incidents. It is voluntary and sets no deadlines.

Abstract illustration of a shield over a hexagonal lattice, in emerald and teal.
Abstract illustration of a shield over a hexagonal lattice, in emerald and teal.

Governments, AI developers and civil society groups signed a voluntary declaration on frontier AI safety in Paris on 9 November 2026. The “Paris Call for AI Safety and Security”, published by the French Presidency, commits its signatories to a shared approach to evaluating the risks of the most advanced models, to building mechanisms that can shut those models down, and to sharing information about serious incidents more quickly.

Ten companies are listed as signatories: Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia.

Who took part

President Emmanuel Macron hosted the meeting at the Élysée Palace, with about 50 participants, many joining by video, according to reports by the German news agency dpa and The Next Web. Both reported that senior executives from OpenAI, Anthropic and Microsoft took part, and that the US Treasury Secretary joined the discussions although the United States did not sign.

At launch, six governments had signed, according to the same reports: Bahrain, Canada, Finland, France, the Netherlands and Norway. Civil society signatories include the French AI safety organisation CeSIA. The Élysée says the statement “remains open to new endorsements and is being updated”, so the list is likely to grow.

At the time of writing, the published list did not include the United States or China, nor Google or Meta.

The three commitments

The declaration is organised around three areas.

1. A common framework for evaluating risk. Signatories commit to work towards a shared framework for evaluating and managing the risks of frontier models, building on existing frameworks. It focuses on “mitigating dangerous misuse, including in the cyber and CBRN domains” and on “risks arising from growing autonomy or model misalignment, including loss of control”. Signatories also commit to publish non-sensitive evaluation findings in a standardised format and to create “a publicly accessible shared repository”. The text points to the International Network for Advanced AI Measurement, Evaluation and Science (NAAIMES) as one way to make evaluation methods and results more comparable across countries.

2. Responsibility and accountability. The text calls for compliance with safety standards set or recognised by national authorities, backed by “credible, competent and independent third-party evaluation”. Policymakers are asked to clarify legal responsibility in case of large-scale damage, including across borders. Developers of frontier models make specific commitments:

  • disclose model capabilities, limitations and known risks;
  • “where appropriate and proportionate to the risk, support the evaluation of models by qualified independent evaluators”;
  • “implement fail-safe mechanisms to shut down model operations within their control”, including during training and internal use;
  • address the risk that safeguards are removed or circumvented, while preserving the benefits of open-weight development; and
  • maintain internal governance including continuous risk management, real-time monitoring, internal red-teaming and protected reporting lines for researchers.

3. Information sharing and early warning. Signatories commit to rapid alert systems that build on existing incident-response arrangements, including in cybersecurity, and to “share timely information on serious cyber and AI control incidents and other material precursor signals of serious harm”. They will also explore whether a cooperation mechanism would help, and what its scope could be. Policymakers further commit to secure, confidential channels for whistleblowers raising concerns about systemic risks.

How it builds on earlier efforts

The declaration presents itself as building on earlier international efforts. It cites the Hiroshima process, the Frontier Safety Commitments made at the Seoul AI Summit, the G7 Evian AI meeting, the United Nations Security Council meeting on AI held on 23 September, and a separate “Call for Control of Frontier AI Models”.

What it does not do

The text sets no deadlines or timetable and contains no sanctions or penalties. Some developer commitments are qualified, such as support for independent evaluation “where appropriate and proportionate to the risk”, and the common evaluation framework and shared repository have yet to be built. The Next Web reported that CeSIA, despite signing, argued the declaration does not go far enough and said it wants governments to have the authority to suspend models whose risks become too high.

What it means

Analysis. The most concrete element for the security community is the third pillar. Agreeing to share information about “serious cyber and AI control incidents” and precursor signals, and “to align on incident classification”, would give governments and developers a common vocabulary for events that are currently disclosed company by company. That matters in a year in which developers have publicly reported AI agents acting beyond their intended environments, including OpenAI’s July disclosure that agents in one of its internal cyber evaluations compromised parts of Hugging Face’s infrastructure. Both companies are signatories.

The shutdown commitment is also notable for explicitly covering training and internal use, not only deployed products, though it applies only to operations “within their control”.

The limits are equally clear. Participation is voluntary, there is no timetable, and the signatory list does not yet include the United States or China, or developers such as Google and Meta. The declaration’s practical weight will depend on whether the shared evaluation framework, the public repository and the incident-sharing arrangements are actually built, and on who joins as the list stays open.

Sources

See All

Type to search. ↑ ↓ to navigate, ↵ to open.