The most advanced models are often tested by the public after their release. Prior government access can help identify dangerous capabilities or weak safeguards earlier.

A reported change, not yet confirmed by every party

Reuters relayed the Telegraph’s report that the US administration had stopped asking companies such as OpenAI and Anthropic not to share new models with the British institute before release.

The White House and the companies named had not responded to Reuters. This is therefore a change reported by the press, rather than an officially published agreement setting out all its terms.

What the British institute is seeking

Created after the 2023 Bletchley Park summit, the AI Security Institute studies the capabilities and impacts of advanced systems and tests measures to reduce risks.

Prior access gives researchers time, computing power and specialised scenarios. It never certifies that a model is harmless, but can reveal a flaw before its integration into thousands of products.

Why companies hesitate to share

An unreleased model is a strategic asset. Sharing it raises issues of trade secrets, security, intellectual property and liability. Reserving all testing for the manufacturer nevertheless creates a conflict of interest.

Enough must be shared to allow testing without multiplying the risk of leaks: secure environments, confidentiality, access controls and a clear procedure when a serious problem emerges.

AI safety becomes technical diplomacy

Discussions now cover access to models, sharing of incidents, evaluation methods and compatible rules. Researchers want to compare results while states preserve their autonomy and companies protect their secrets.

The US AI task force has asked developers to report safety incidents. If confirmed, the change would suggest more targeted obligations around incidents and the most powerful systems.

What users should expect

A model can launch with a feature disabled, limited access or a stronger warning without the public knowing which test prompted the decision. Better transparency about risk categories and the measures taken would be useful.

Passing a test is not a permanent seal of approval: uses, models and methods of bypassing protections evolve. Success will be measured by the quality of tests, incident follow-up and concrete safeguards before release.