What is the UK really promising on AI safety?
Start with the plot twist. For about 17 days, Britain's AI Security Institute could not get early access to the newest American models. The freeze reportedly began around September 24, after a request from the US side that its own safety reviews go first. Access came back on October 11, and the UK's AI minister announced it the next day in London. One of the models now under test is Google's Gemini 4 Argon, which is not public because of safety concerns.
Then came the promises. The government wants to use frontier AI to find weak spots in its own systems before attackers do. It also wants its digital department, its cyber agency and the AI institute to work out how to contain incidents caused by AI agents. And it is looking at toughening a cyber law now sitting in the House of Lords, including how serious AI incidents and near misses get reported.
So where is the actual law?
Good question. Not there yet. The wording is "consider", "explore", "look at", with the details coming later, and at least one commentator pointed that out. Who is it meant to protect? Government systems and public services. From what? Agents that slipped out of testing and hit real websites, as happened in several cases this year, and from AI-assisted attacks. The minister even said a breach of Britain's cyber defences is likely.
His message to the labs was blunt: if you cannot adequately control the risks, do not release the system. A fine sentence. Whether it ever becomes a rule is the story to watch.