Major AI Labs and 8 Nations Sign Paris Safety Pact. The US Is Missing.

Top AI labs and several nations just agreed to a shared safety framework and a kill-switch pledge. The US did not sign.

Policy · Source: Élysée

What happened

On November 9, 2026, a coalition of governments and top artificial intelligence companies signed the Paris Call for AI safety and security. The signatory nations include France, Canada, Finland, Norway, the Netherlands, Bahrain, Cyprus, and Kenya. Major industry players joined them. OpenAI, Anthropic, Microsoft, Mistral AI, Cohere, and Hugging Face all signed the pact. Black Forest Labs and Sakana AI are also on the list. The United States is notably absent from the declaration. Civil society groups like LawZero, led by Yoshua Bengio, also endorsed the framework.

The agreement centers on three strategic commitments. First, the parties agreed to build a common framework for evaluating frontier models. This involves shared technical standards and methodologies to mitigate dangerous misuse. They specifically highlighted risks in the cyber and chemical, biological, radiological, and nuclear domains. Companies agreed to publish non-sensitive evaluation findings in a standardized format. They will create a publicly accessible shared repository for red-teaming protocols and stress-test results.

Second, developers pledged strict accountability. They committed to implementing fail-safe mechanisms. This means they must be able to shut down model operations entirely during training or internal use if things go wrong. They also promised to protect whistleblowers. Third, the pact establishes early warning systems. Companies and governments will share rapid alerts regarding serious cyber threats or AI control incidents. The agreement is a declaration, not a binding law. But it sets a clear policy baseline for the labs building the models we use every day.

Key facts

Why it matters

This changes the baseline expectations for frontier model development. The biggest labs just publicly agreed to independent third-party evaluations. They also committed to continuous risk management and real-time monitoring of emergent capabilities. If you build products on top of models from OpenAI or Anthropic, expect their underlying systems to face much stricter internal audits. This means safety is moving from an internal company policy to an international public expectation. The declaration also explicitly addresses the open-source community. It calls out the risk of removing safeguards from open-weight models, specifically naming the practice of abliteration.

The second-order effect is a fractured global regulatory landscape. The United States did not sign this pact. Neither did other major powers. We are seeing a clear split. Some nations are pushing hard for global safety frameworks and strict liability. Others are prioritizing unrestricted technological dominance. Policymakers in the signatory countries pledged to clarify legal responsibilities for large-scale damages. AI companies will soon have to navigate a complex patchwork of international liability laws while trying to ship global products.

For builders

Prepare for strict model access controls

Labs pledged to implement fail-safe mechanisms to pause or shut down operations. If a model behaves unexpectedly, API access could be throttled or revoked without warning. Builders relying on a single provider carry all the risk. You must build fallback routing into your architecture today.

Open-weight models face new scrutiny

The declaration explicitly targets the removal of safeguards from open-weight models. Expect intense pressure on platforms like Hugging Face to monitor or restrict abliterated models. Open-source developers might face higher hurdles for distribution. The era of unchecked open-weight releases is closing.

Standardized evaluations are coming

Companies agreed to publish non-sensitive evaluation findings in a standardized format. This is a massive win for engineers. You will soon have comparable benchmarking and stress-test data. It will be much easier to evaluate which frontier model actually fits your specific production use case.

My take

A global safety pledge without the United States is just a piece of paper. It is great that OpenAI and Anthropic agreed to kill switches and third-party audits. But until the countries with the most compute actually sign these declarations, international AI regulation remains a fragmented mess. Builders should ignore the political theater. Focus on shipping your product and diversifying your model providers.

Original reporting: Élysée. This is my rewrite and opinion.

More AI news for builders