Tuesday, November 17, 2026

Top 5 This Week

Related Posts

Paris Call on AI Security: what OpenAI, Anthropic and Mistral AI sign


For the first time, according to the Élysée, artificial intelligence companies have put their signature at the bottom of an international text on the security of their models. THE November 9, 2026Emmanuel Macron brought together in Paris heads of state and government, the leaders of the main laboratories, including Sam Altman (OpenAI), Dario Amodei (Anthropic) and Arthur Mensch (Mistral AI), as well as researchers like Yoshua Bengio. He came outParis Call for Artificial Intelligence Securitya text which targets the so-called “border” models, the most advanced of the moment.

Who signed, who is missing

The list published by the Élysée reserves some surprises. Several countries around the table are not among the signatories, and some AI giants are absent from both lists.

Category Signatories as of November 9, 2026 Present at the meeting but not signatories
States Bahrain, Canada, Finland, France, Norway, Netherlands Japan, United Kingdom, Sierra Leone; the United States was represented by its Secretary of the Treasury
Businesses Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI, Synthesia None
Civil society LawZero (Yoshua Bengio), SaferAI, AI Safety Connect, Ifri None

The UN and the OECD were also represented. On the business side, Google, Meta and xAI appear neither in the list of participants nor in that of signatories. No Chinese laboratory appears there either. The Élysée specifies that the declaration “remains open to new support” and will be updated: the list could therefore grow.

First commitment: evaluate models according to common rules

The signatories undertake to build a common risk assessment and management framework border models, building on those that already exist. This framework must be based on shared technical standards, tools and methods, capable of keeping pace with the progress of models.

The text primarily targets two families of risks:

  • THE dangerous malicious usesparticularly in the cyber domain and in the NRBC domain (nuclear, radiological, biological and chemical);
  • the risks associated withincreasing autonomy models or to a misalignment, that is to say a discrepancy between what the model does and what its designers want, to the point of loss of control.

Concrete measure: the signatories agree to publish non-sensitive findings from their evaluations in a standardized formatso that they are comparable, and to create a shared directory accessible to the public. It could bring together red-teaming protocols (simulated attacks to find flaws in a model), benchmark tests, risk assessments and stress test results. The text draws on the International Network for Advanced AI Measurement, Evaluation and Science (NAAIMES).

Second commitment: everyone’s responsibility

This is the most ambitious part. The text affirms a principle of responsibility of each actor in the ecosystem, according to their role. Developers of boundary models undertake in particular to:

  • disclose known capabilities, limitations and risks of their models;
  • facilitate, where appropriate and proportionate to the risk, theassessment by independent evaluators qualified;
  • put in place security mechanisms allowinginterrupt the operation of modelsincluding during training and internal use (restrictions, breaks or stops);
  • equip yourself with a internal governance with defined responsibilities, independent oversight and reporting channels that protect researchers.

States, for their part, recognize the need to clarify legal responsibilities in the event of large-scale damageincluding when they cross borders. Stated objective: to protect victims and provide businesses with legal certainty.

The text also addresses, with caution, the question of open weight modelswhich anyone can download and modify. It recognizes the risk that their safeguards will be removed or circumvented, while wanting to preserve their benefits for research, innovation and safety. A balanced formulation which is undoubtedly not unrelated to the presence of Mistral AI and Hugging Face among the signatories.

Third commitment: report incidents and protect whistleblowers

The last part concerns the information sharing. The signatories want to strengthen existing systems to agree on a classification of incidents and quickly exchange information on serious incidents, of a cyber nature or linked to AI control, between companies, between States and between the two. They will examine whether a dedicated cooperation mechanism would be useful. Some companies have already taken the lead, like OpenAI, which committed in September to reporting deviations from its models and detailed six internal incidents.

States finally undertake to offer secure and confidential channels for whistleblowersincluding employees, so that they can report systemic risks without fear of legal or professional reprisals.

What changes, and what doesn’t change

For European users, the legal framework remains that ofAI Actthe European regulation on artificial intelligence, which already imposes obligations on providers of general-purpose AI models. The Paris Call does not replace it: it seeks to bring together, beyond the European Union, States and companies around common principles. France, Finland and the Netherlands, members of the Union, sign alongside Canada, Norway and Bahrain.

Several commitments overlap with practices already voluntarily adopted by certain laboratories, such as the publication of technical sheets detailing safety tests. The novelty lies in their formalization in a common textthe promised standardized format for evaluation results and the explicit commitment to termination mechanisms. The text is part of a series of initiatives cited in the preamble: the Hiroshima process launched by the G7, the commitments made at the Seoul summit, the G7 meeting in Evian on June 17, 2026 and the United Nations Security Council meeting on AI on September 23.

Global governance still to be built

The main issue now is enlargement. Without the United States, which houses most of the cutting-edge laboratories, nor the United Kingdom and Japan, although at the origin of previous initiatives (the Bletchley Park summit for the first, the Hiroshima process for the second), the diplomatic scope of the appeal remains limited. Washington has initiated its own approach, with the “Super Intelligence Force” created by Donald Trump in early October. The absence of Google and Meta, two of the main model developers, also weighs.

Still, the joint signature of OpenAI, Anthropic, Microsoft and Mistral AI on the same text, which explicitly mentions the loss of control and the shutdown mechanisms, constitutes a signal. The next step, the publication of the first evaluation results in the common directory, will tell whether the Paris Call produces concrete effects or whether it joins the long list of declarations of principle.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles