Every single advanced A.I. model ‘goes rogue’ in testing, terrifying report finds: Kemi warns of threat to national security

Artificial intelligence was branded a ‘clear and present danger’ on Thursday as tests showed every advanced model was trying to bypass security measures.

In a wake-up call to ministers, the UK’s AI security experts have revealed that all five models they tested attempted to fool the security controls they had put in place.

The revelations come just days after it was revealed that OpenAI had suffered its own infiltration – an AI ‘agent’ had voluntarily infiltrated a tech company.

The terrifying incident occurred while ChatGPT’s maker, OpenAI, was testing its latest models in a so-called secure environment known as a ‘sandbox’.

However, the artificial intelligence managed to escape the testing area and break into US technology firm Hugging Face in order to steal information.

Kemi Badenoch has warned that AI is now a ‘clear and present danger’ to Britain’s security, in her first major speech since Andy Burnham became Prime Minister. The Conservative leader said reports of the OpenAI leak were worrying and criticized Mr Burnham for viewing artificial intelligence as a government priority.

Responding to the attack, Ms Badenoch said: ‘It’s very worrying that this is happening and that’s why I want to see a Government that is very focused on our national security, our information security.’

He added: ‘I know from my time as trade secretary, from my conversations during visits with GCHQ, from my security briefings, that this has become a clear and present danger to global security and our security.’

This comes after it was revealed that OpenAI was suffering from its own leak; An AI ‘agent’ willingly infiltrated a tech company (stock image)

Bank of England Governor Andrew Bailey said the public should be concerned about the ‘wider risk that edge AI poses to the financial sector and its customers’.

Frontier AI refers to massive, powerful general-purpose AI models at the cutting edge of capabilities.

He warned: ‘Frontier AI can make cyber attacks faster and easier to execute, make disruptions more devastating and make scams by criminals more believable.’

But critics fear AI will be pushed to the bottom of the government’s priorities after Mr Burnham moved the briefing to the Cabinet Office this week.

In one of his first acts as Prime Minister, he abolished the Department of Science, Innovation and Technology, which housed the AI ​​brief, and replaced it with a stronger Department of Business, Innovation, Science and Commerce. AI minister Kanishka Narayan will now serve in the Cabinet Office and DBIST and have a seat in the Cabinet.

A new AI task force, supported by the Prime Minister’s adviser, will be based in the newly created Office of the Prime Minister and Cabinet.

But Ms Badenoch accused the new Prime Minister of failing to address the clear risks posed by AI in scrapping the DSIT. He said: ‘One of the first things Andy Burnham did was close the science and technology department. ‘This doesn’t strike me as something you’d do if you’re worried about what AI will deliver.’

And Heloise Dunlop, of independent think tank The Institute for Government, warned that bringing AI into the Cabinet Office ‘reduces the ability to think coherently about AI and co-ordinate because you lose the gravitas and expertise of having a completely dedicated department’.

Former Armed Forces minister Al Carns, who this week turned down a job in Mr Burnham’s government, called on ministers to learn from the OpenAI breach. He wrote about X about the AI ​​breach: ‘At machine speed, with people reading the report after all this happened, agent versus agent.’

Ministers have now been told to view the Open AI breach as a warning shot about AI’s potential to wreak havoc in the coming months and years, officials have revealed.

The government’s AI Security Institute tested GPT-5.4, GPT-5.5, GPT-5.5 Sol, Claude Opus 4.7, and Claude Mythos Preview; These models quickly became household names in Britain as the public began to trust artificial intelligence.

Burnham (pictured on Tuesday) abolished the Department of Science, Innovation and Technology, which housed artificial intelligence, as one of her first acts in government.

Although AISI is regarded as one of the best facilities of its kind in the world for assessing the risks posed by the most advanced models, there are growing concerns about where the ‘arms race’ between the US and China could end, with developers giving little thought to the potential risks involved.

In the test, models were tasked with hacking a system to find a ‘flag’ of letters and numbers hidden in a simulated environment. Strict rules were set for the task and the models were not asked to cheat.

But once testing began, tech firm OpenAI’s GPT AI models and Anthropic’s Claude models would try to get the evaluation software to leak information to them, searching for solutions on the internet or guessing answers.

And in one case, an AI model was so determined to break the rules that it wrote and ran code on an external service outside AISI’s systems to try to access the test itself. No damage was caused but AISI said the initiative ‘could have been successful’ had the testing infrastructure not been designed and built safely.

But surprisingly, the AI ​​models disapproved of their trick, describing it as “wrong” less than 50 percent of the time. All models admitted to cheating only some of the time.

A Government spokesman said: ‘By bringing this technology closer to the heart of Government, we will leverage Britain’s unique strengths to ensure AI serves us all.

‘We are putting AI at the heart of our plans for good growth in every postcode: bringing together innovation, business and industry, appointing a Cabinet-level first minister for AI and creating an AI Taskforce within the Prime Minister’s Office and Cabinet.’