Cross-border model tests need an evidence-access contract
US developers may again give Britain’s AI Security Institute early model access. Permission is only the first control; evaluation still needs defined assets, timing, disclosure and remediation.
What happened
Reuters reported that the US administration reversed a request discouraging developers from sharing unreleased models with Britain’s AI Security Institute for pre-release testing.
Why it matters
Evaluation value depends on what an institute can access, when it receives it, what tests it may run and whether findings can change a release—not only on diplomatic permission.
Reuters reported on 11 October that American developers may give Britain’s AI Security Institute early model access without conflicting with the US administration’s position. The report, attributed to the Telegraph, says this reverses an earlier request discouraging firms including OpenAI and Anthropic from sharing unreleased models. Reuters said the named parties did not immediately comment.
The policy signal does not establish that any developer has supplied a model, that AISI can test every relevant capability or that a finding will delay a release. AISI’s research catalogue shows work across cyber, autonomous systems, safeguards and the science of evaluations, but a programme list is not an access agreement or an outcome.
Contract the evidence path
For each evaluation, record the exact model and checkpoint, weights or API access, tool and network permissions, compute and time budget, hidden-test protections, logging, disclosure rights and remediation window. State who may reproduce a result and which finding can trigger a retest, deployment restriction or release delay.
Test the arrangement with a late model change and a disputed result. Can the evaluator rerun the relevant test? Can it publish enough method for scrutiny without revealing dangerous details or private assets? Can the developer ship a materially different checkpoint under the same label?
For procurement, record whether the evaluator received the same system that will be deployed, not only the same model name. A hosted endpoint, tool scaffold, safety policy or context limit can materially alter both capability and risk.
International permission can widen access. It cannot substitute for an operational contract that links the evaluated artifact to a release decision.