Forty Governments Agree to Share AI Safety Test Results

Three years after the first intergovernmental summit on advanced artificial intelligence, the diplomacy has produced something more concrete than a declaration. In this demonstration scenario, forty governments agreed to share the results of safety evaluations conducted by their national AI institutes before advanced models are released.

The agreement builds on the network of safety institutes that governments committed to in Seoul in 2024. Under its terms, a model tested by one institute would have its evaluation summary made available to the others, and signatories commit to a common format for reporting dangerous capabilities.

What is not in the text

The agreement does not require testing, does not set thresholds for what counts as unsafe, and does not say whether results will ever be made public. Each of those gaps was the subject of negotiation, and each was left open to secure signatures.

More consequential is who is absent. Two of the largest AI-producing countries attended the talks but did not sign, citing concerns about the protection of commercially sensitive information. Without them, the shared pool of results will cover a minority of the most capable models.

The record so far

The path to this point runs through Bletchley Park in 2023, where governments first acknowledged frontier risks together; Seoul in 2024, where companies made voluntary commitments; and Paris in 2025, where the agenda broadened to investment and inclusion. The European Union's AI Act, in force since August 2024, remains the only binding regime among the major economies.

Supporters of the new agreement argue that information sharing is the step most likely to survive changes of government because it costs little and constrains nothing. Critics make the same observation in reverse.