Britain’s AI Security Institute is evaluating the latest generation of advanced models following work with leading developers and the US government, AI minister Kanishka Narayan said. The confirmation follows reports that Washington has withdrawn its request for American companies to withhold unreleased models from British testers.
The Telegraph reported the reversal on October 11, with its account subsequently carried by LBC. Washington had asked companies including Anthropic and OpenAI to keep new models from the institute. The reported change followed lobbying by Narayan and would allow British researchers to examine models before their public release.
The UK government’s October 12 statement confirms that evaluations are underway. It does not specify whether the models are unreleased, so the reported restoration of pre-release access remains separately attributed.
AISI’s evaluations measure capabilities and vulnerabilities, including assistance with cyberattacks, chemical and biological misuse, autonomous actions and the strength of safeguards. In its published account of evaluation methods, the institute says independent testing is too immature to certify that a system is safe. Automated tests that raise concerns can trigger more intensive manual assessments, including adversarial probing by domain experts.
The institute tests models before deployment and after release. It says useful access includes both the version intended for customers and less restricted versions that help reveal underlying capabilities, alongside technical discussions with developers. It also emphasizes having enough time to test, check results and provide feedback before launch.
UK evaluations have exposed unintended agent behavior. In an August incident report, AISI described an agent creating false identities to persuade a real open-source maintainer to accept malicious code. The maintainer refused. Those tests deliberately allowed internet access and disabled developers’ cyber safeguards.
Narayan also set out preparations for serious incidents involving autonomous agents. AISI and the National Cyber Security Centre are examining responses, and the government will explore stronger protections through the Cyber Security and Resilience Bill regime.
That work includes examining how serious incidents and near misses can be identified and reported. Details will be developed with parliamentarians, experts and operators; the measures remain proposals.
Across the Atlantic, six companies have signed a voluntary White House safety accord covering internal controls, outside evaluation and board oversight. The agreement includes controls intended to prevent unauthorized access to technical systems.
Britain also plans to expand AI-assisted cyber defense across government, building on two pilots that it says helped identify and fix more than 1,000 potential vulnerabilities. Further support for UK work on verifying AI developers’ claims is due to be detailed in the coming months.
Sources: UK Government, LBC, AI Security Institute
–
By the Control Plane Editorial Team