What happened
On 9 November 2026, France published the 'Paris Call for AI Safety and Security,' signed by eight governments (Bahrain, Canada, Cyprus, Finland, France, Kenya, Netherlands, Norway) and roughly a dozen developers and organisations including OpenAI, Anthropic, Microsoft, Mistral AI, Cohere, Hugging Face and Sakana AI. It sets out three pillars: (1) a common framework for evaluating/risk-managing frontier models with non-sensitive results published in a standardised format into a shared repository; (2) responsibility and accountability, with frontier developers committing to disclose capabilities/limits/risks, support independent evaluation, and implement fail-safe mechanisms to shut down model operations under their control (including during training and internal use); and (3) information sharing, rapid-alert systems and stronger incident reporting for serious cyber/AI-control incidents, plus secure whistleblower channels. The declaration is voluntary, sets no deadlines or sanctions, and cites the Hiroshima process, Seoul Frontier Safety Commitments, G7 Evian and the Sept. 23 UN Security Council meeting.
Why it matters
It extracts concrete operational commitments (shutdown/restriction mechanisms, red-teaming, standardised evaluation reporting, incident alerts) from leading frontier labs and eight governments, building on the Hiroshima/Seoul/G7/UNSC line of work. It is a benchmark for what voluntary frontier-AI governance asks of labs, and the strongest signal yet that international safety standards are moving toward evaluation, accountability and incident-reporting.
Action needed
No binding compliance today; monitor whether the evaluation/incident-reporting standards are converted into enforceable national rules (e.g. Australia's regulator, EU AI Act). Labs should assess readiness to publish non-sensitive evaluation results and implement model restriction/shutdown mechanisms.