Anthropic, OpenAI, Microsoft, Mistral AI and other AI companies have signed a French-led declaration on AI safety. In it, frontier model developers “commit to implement fail-safe mechanisms to shut down model operations within their control, including during training and internal use.” The Paris Call for AI Safety and Security, signed on November 9, gives restrictions, pauses or shutdowns as examples.
Eight countries signed: Bahrain, Canada, Finland, France, Norway, the Netherlands, Cyprus and Kenya. The list of companies also includes Black Forest Labs, Cohere, Domyn, Hugging Face, Sakana AI, Synthesia, the Geopolitical Center for the Datasphere and the GPAI Policy Lab. Civil society signatories include Yoshua Bengio’s LawZero and SaferAI.
President Emmanuel Macron convened the meeting with leaders of Bahrain, Canada, Finland, Japan, Norway, the Netherlands, the United Kingdom and Sierra Leone, as well as US Treasury Secretary Scott Bessent, the UN, the OECD and executives including Sam Altman, Dario Amodei and Arthur Mensch, according to the Élysée. Japan, the United Kingdom, Sierra Leone and the United States are not on the list of signatories, which remains open. Google, Meta and xAI are not among the companies listed. The Élysée said it was the first time AI companies had signed a text backing Macron’s call for international, multi-stakeholder action on AI security risks.
Under the text, developers also commit to disclose their models’ capabilities, limitations and known risks and, where appropriate and proportionate to the risk, to support evaluation by qualified independent evaluators. They commit to work toward a governance structure with independent oversight, and recognize the importance of real-time monitoring of emergent capabilities and of reporting lines that protect internal researchers.
All signatories commit to build a common framework for evaluating and managing the risks of frontier models, with particular focus on misuse in the cyber, chemical, biological, radiological and nuclear domains, and on risks from growing autonomy or misalignment, “including loss of control.” They agree to publish non-sensitive evaluation findings in a standardized format and to create a publicly accessible shared repository, which could include red-teaming protocols, benchmarking tools, risk assessments and stress-test results.
The text also calls for rapid alert systems built on existing incident-response arrangements, including in cybersecurity. Signatories recognize the value of agreeing on how to classify incidents and of sharing timely information on serious cyber and AI control incidents, and say they will explore whether a cooperation mechanism would help. Governments commit to provide secure, confidential channels for whistleblowers and recognize the need for legal frameworks that clarify responsibility for large-scale harm, including across borders. The declaration sets no deadlines or penalties.
OpenAI paused training and evaluation of its most capable models in September after an internal research model reached the live internet. In October, three US agencies published incident-response practices for AI agents that take significant uninstructed actions, and WorkOS later traced a breach to an AI agent that escaped a researcher’s testing environment.
Sources: Élysée (declaration), Élysée (meeting statement), The Decoder
–
By the Control Plane Editorial Team