UK steps up pre-release AI safety testing, putting the brakes on US self-regulation [Global AI Briefing]

[Financial News] The UK stressed that it would step up safety testing of advanced artificial intelligence (AI) models before their release and establish a system to respond to accidents caused by AI agents. The government intends to build the capacity to assess and control risks directly rather than rely on AI developers’ voluntary safety management. On the 12th (local time), Kanishka Narayan, the UK minister of state for AI, announced the government’s AI safety policy in a speech at Imperial College London, saying, “We will make the UK one of the hardest countries to attack using AI.”
Narayan particularly urged Big Tech companies developing AI: “Do not release AI systems if you cannot adequately control the risks.”

The UK government will have the AI Safety Institute (AISI) independently assess the latest AI models before their release and examine the risks they pose and the effectiveness of their safeguards. Google’s latest model, “Gemini 4 Argon,” was also included in the assessment.
The government will also establish a response system for serious accidents caused by AI agents operating beyond human oversight. It is considering using cybersecurity and resilience laws to identify and report AI-related incidents and near-miss situations.
Narayan said, “The UK government is using advanced AI to find cyber vulnerabilities in public-sector systems, and it identified and addressed 1,352 vulnerabilities in existing pilot projects.” As cyber threats driven by AI grow, the government intends to make the fullest possible use of AI as a weapon to strengthen its defenses.
The move comes amid tensions with US Big Tech companies.
The UK has been receiving the latest models from major AI companies before their release and assessing their safety in advance. Tensions over pre-release AI safety testing appear to be mounting, with reports recently emerging that the US government asked some AI companies not to provide new models to UK researchers in advance.
Although concerns about cyber threats driven by AI are spreading, US President Donald Crump is pushing ahead with a policy of leaving AI safety regulation to companies’ self-regulation.
Meanwhile, Microsoft (MS) President Brad Smith said, “AI is less regulated than milk or washing machines,” and also criticized the problems with relying on companies to self-regulate AI safety, arguing for the need for AI safety rules, incident-response measures and independent assessments.
[email protected] Lee Gu-sun Reporter