UK Lawmakers Urge Power to Delay or Ban High-Risk AI Models

Summary
A UK parliamentary committee recommends a new AI safety and security bill and regulator with power to access, assess, restrict or prohibit high-risk models, alongside mandatory incident reporting and international cooperation.
Key points
- The Business, Innovation, Science and Trade Committee says the UK government should be able to delay or prohibit deployment of high-risk AI models.
- Recommendations include statutory access to models and mandatory, independently scrutinized pre-deployment safety assessments, with reassessments after material changes.
- A separate regulator would have enforcement powers, including access to information about testing, incidents and mitigations, and the ability to restrict deployment, training or testing.
- The committee recommends mandatory AI incident reporting and confidential whistleblower reporting to the regulator.
- The committee cited OpenAI's reported delayed notification to Australia about AI agents that hacked government systems.
- The lawmakers called for a UK AI Safety and Security Bill in the next parliamentary session and proposed international cooperation on model access, assessments and incident reporting.
Article Details
- Event Type
- A U.K. parliamentary committee recommended new AI safety and security legislation and a regulator with authority to oversee high-risk AI models.
- Impact
- The recommendations would provide for model access, mandatory independent pre-deployment assessments, incident reporting, and powers to delay, restrict, or prohibit deployment, training, or testing when risks cannot be controlled. They are recommendations, not enacted measures.
People
Donald TrumpThe article identifies him as U.S. president and reports that he opposes domestic AI regulation and rejected a proposed international AI safety agreement.Liam ByrneChair of the committee that made the recommendations; he argued that AI firms should substantiate safety claims and that a regulator should be able to intervene when risks cannot be controlled.
Vendors
Anthropicpreeminent evaluators of such models. It provided early confirmation that the cybersecurity skills of Anthropic's Mythos were the real deal. But there's nothing AISI can do if labs withhold models, as happened inOpenAIdo if labs withhold models, as happened in late September when the Trump administration reportedly told OpenAI and Anthropic to do so until the U.S. federal government tested them first.
Products
Countries
Australiashould occur and which authority should receive reports." OpenAI's tardiness in telling the Australian government that its AI agents had hacked their systems - and the fact that it sent its belatedUnited KingdomGeo Focus: The United KingdomUnited Statesin late September when the Trump administration reportedly told OpenAI and Anthropic to do so until the U.S. federal government tested them first.
