Eight countries and top AI labs agree to model shutdown controls
The Paris text makes labs agree to shutdown controls for their models, but it sets no deadline and no sanction.
On 9 November 2026, eight countries, AI companies and civil groups agreed to the Paris Call for AI Safety and Security. The French president held the meeting on the internet. The leaders of OpenAI and Anthropic were in it, with the US Treasury Secretary, but Germany did not agree. The text is a political promise and not a legal rule.
Agreed controls
The companies agree to make shutdown controls for the models they operate. They also agree to tell about known risks and to help external researchers do audits.
Non-sensitive test results go out in one standard format. The text includes a set of tests for strong models and a procedure to tell each other about important cybersecurity incidents.
The proof
The source gives no test results. The text sets no deadline and no sanction. The source gives no cause why Germany did not agree.
Open questions
The standards and the procedures to work together are not made at this time. The source does not tell how to check a shutdown control. At this time, the text is a political promise only.
Insight: Shutdown controls become a common term
On 17 October, three US agencies told companies to write policies on when to stop a model. In October, Anthropic and OpenAI sent California two plans to stop a dangerous model. The Paris text also names shutdown controls. This reads as a race to set the test for a shutdown control, because the first accepted test can become the rule. Vendors must show the controls they hold.
Sources
Posted
