A young programmer in Prague trains a new AI model in her spare time. A hospital in Brno tests an AI diagnostic tool. A start-up in Warsaw builds customer-service chatbots for banks across Europe. All of them — whether they know it or not — depend on someone, somewhere, checking whether the AI they use is actually safe. Right now, the country doing the most to fill that role is the United Kingdom. But its lead comes with a catch: the institute doing the checking has no power to stop anyone from releasing a dangerous model.
The institute that tests but cannot ban
When Anthropic decided that its latest model, Claude Mythos, was too dangerous to release freely, it gave pre-release access to only one non-American organisation: the UK's AI Security Institute (AISI). In April 2026, the London-based team published its evaluation, warning that Mythos could hack into critical infrastructure that was previously considered secure. The model completed a 32-step simulated corporate network attack from start to finish — something no AI had done before. Tasks that would take a human expert 20 hours, Mythos handled autonomously in minutes. The Mythos case is the most dramatic example, but it is not the only one. Before OpenAI released GPT-5, AISI found more than a dozen vulnerabilities that could have allowed users to develop biological weapons. All were fixed before the model went public. The institute has now evaluated more than 30 frontier models and, according to its own tracking, has uncovered safety gaps in every leading AI system it has tested. This is independent evaluation that actually matters — AI companies marking their own homework is a well-known problem, and AISI provides an outside check that governments and the public can trust. The institute publishes its testing tools on an open-source platform called Inspect, so researchers worldwide can replicate its methods. But here is the part that should worry everyone who uses AI: AISI has no legal authority whatsoever. It cannot force an AI company to submit a model for testing before release. It cannot block a dangerous model from going to market. It cannot impose fines. It is, in legal terms, a research body inside the civil service — not a regulator. This gap is not theoretical. In April 2026, AISI evaluated OpenAI's GPT-5.5 and found a universal jailbreak that defeated the model's cyber safeguards across every single malicious query it tested. AISI flagged the issue before release. The model was launched publicly anyway, before the institute could confirm whether the vulnerability had been fixed. Nothing happened. No fine. No recall. No suspension. Because AISI had no power to stop it.Who checks the checkers?
The Ada Lovelace Institute, an independent research body in London, published a comprehensive analysis of AISI on 13 July 2026. The verdict? "World-leading work" — but with serious structural limits. AISI focuses narrowly on catastrophic and national security risks: cyber attacks, biological weapons, loss of control. That is important work, but it leaves huge areas untouched. The institute does not test for bias and discrimination in AI hiring tools, or for the generation of non-consensual intimate images, or for the impact of AI companion apps on teenagers' mental health. These are precisely the harms that regular people encounter every day. The Ada Lovelace analysis also points out another uncomfortable fact: AISI's access to models depends entirely on maintaining warm relationships with the same companies it is supposed to scrutinise. The UK government signs memoranda of understanding with OpenAI, Anthropic, Google DeepMind, and others — partnerships that bundle evaluation access with broader government collaboration on public services and research funding. "How independently can a body assess models built by companies it is simultaneously partnered with?" the report asks. The testing windows are also short. When Anthropic released Claude Sonnet 4.5, AISI had less than a week to evaluate it. By comparison, the EU AI Act's Code of Practice recommends a minimum of 20 days for pre-release evaluation of general-purpose AI.What this means for Europeans
The UK's approach stands in sharp contrast to the European Union's. Where the UK has built technical capacity without legal teeth, the EU has built legal architecture without a single central testing body. The EU AI Act, which entered into force in August 2024 and is now being phased in through 2026–2027, classifies AI systems by risk level and imposes binding requirements on high-risk applications. Companies that break the rules face fines of up to 7% of global annual turnover. But the EU does not have a London-style institute that can independently test the most advanced models before they reach European users. The AI Act requires developers of general-purpose AI to cooperate with the newly created EU AI Office, but the practical testing infrastructure is still being built. The result is an odd transatlantic inversion: the post-Brexit UK has the world's leading government AI testing lab, while the EU has the world's most comprehensive AI law — and neither side has both. For European AI users, this means the safety net is split across the Channel. When you use ChatGPT, Claude, or Gemini from Prague, Paris, or Berlin, the model may have been tested by the British institute — but only if the American company that built it chose to share it. And if the testers found something dangerous, they had no power to prevent it from reaching you.The global patchwork
The UK is not alone in trying to figure this out. A growing number of countries have launched their own safety institutes, but no two operate the same way:| Country / bloc | Has a safety institute | Can compel testing | Can block release | Can impose fines |
|---|---|---|---|---|
| United Kingdom | Yes — AISI (est. 2023) | No | No | No |
| European Union | EU AI Office (est. 2024) | Yes, under AI Act | Yes, for high-risk systems | Yes — up to 7% of global turnover |
| United States | Center for AI Standards (Commerce Dept.) | Voluntary framework | No federal authority | No |
| Japan | Yes — leads G7 AI Safety project | No | No | No |
| Singapore | Yes — AI Verify Foundation | No | No | No |
| China | Yes — established 2025 | Yes | Yes | Yes |
Where the UK got it right
Despite its legal limits, the UK institute's success deserves attention from European policymakers. Three things set it apart. First, political will and funding came early. The institute was born at the Bletchley Park AI Safety Summit in November 2023 and received £66 million (roughly €77 million) in annual funding, plus priority access to over £1.5 billion (€1.75 billion) of government computing power. That allowed it to hire over 100 technical specialists, many poached from Google DeepMind and other top labs — creating one of the largest concentrations of AI-testing expertise in any government worldwide, with salary structures more competitive than the standard civil service. Second, it publishes results openly. The Inspect platform makes testing frameworks available to researchers and governments globally. When AISI finds a vulnerability, the world knows about it — not just a classified briefing in Washington. Third, it built trust with AI companies through technical credibility. Anthropic co-founder Jack Clark has called third-party testing "a really important part of the AI ecosystem," and the fact that companies voluntarily share unreleased models — even for short windows — shows that the institute's reputation opens doors that legal mandates have not.What should change
The Ada Lovelace Institute's recommendation is clear: the UK needs a dedicated AI regulator with statutory powers to compel access for testing, set conditions for public release, and intervene when harms emerge. The EU AI Act already provides much of this framework, but lacks the technical testing muscle that London has built. For European citizens, the practical question is simple: will the people checking whether AI is safe actually be able to stop a dangerous model before it reaches your phone, your workplace, or your child's school? Right now, the answer is: only partially, and it depends on where you live. The combined picture — British testing capacity plus European legal authority — would be far stronger than either side alone. Whether that cooperation materialises in a post-Brexit landscape is one of the most important unanswered questions in AI governance today.Does the EU AI Act already require AI companies to submit models for safety testing?
Yes, for general-purpose AI models that pose systemic risks. The AI Act requires developers to conduct model evaluations, implement risk management, and cooperate with the EU AI Office. The Code of Practice recommends a 20-day minimum evaluation window. However, the EU does not yet have a central testing body comparable to the UK's AISI — the practical testing infrastructure is still being built, and enforcement depends on national authorities in each member state.
If the UK institute finds a dangerous flaw in an AI model, what happens next?
AISI shares its findings with the UK government and, where possible, works directly with the AI company to fix the issue. But it has no power to stop the model's release. The company can — and in the case of GPT-5.5, did — launch the model publicly before the identified vulnerability was resolved. There is no legal mechanism for AISI to block deployment.
What can ordinary Europeans do to protect themselves from unsafe AI?
Stay informed about which models you use and what their safety records look like. The EU AI Act gives you rights: from 2026 onward, high-risk AI systems used in areas like hiring, credit scoring, and law enforcement must meet transparency and safety requirements, and you can file complaints with national data protection authorities if you suspect violations. For consumer AI tools like chatbots and image generators, check the provider's published safety documentation (system cards, model evaluations) before trusting them with sensitive information.
