Access to the latest models

Britain's AI Security Institute (AISI) has been given access to the latest advanced AI models and is working with the US government and developers including Google to assess their safety, AI Minister Kanishka Narayan said on Monday, 12 October. Under voluntary agreements with AI developers, the institute receives access to models before deployment, which lets it assess their capabilities and risks and contribute to emerging international safety standards. The announcement did not specify all the models the institute is evaluating.

“I can confirm that AI scientists at the AI Security Institute are now conducting rigorous evaluations of the latest generation of models, including Gemini 4 Argon, to identify risks early, to test robust safeguards, to help keep people safe.”

Reports in September said the White House had asked OpenAI and Anthropic to withhold new models from British testers pending a US review. In a letter published in September, the AISI said Anthropic had made clear that no organisation outside the US had access to its Mythos 5.1 model. Asked about access to frontier systems, Narayan said the government had a long-standing relationship with the US built on trust. Microsoft president Brad Smith told the BBC, in an interview published on Monday, that Britain was well placed to help lead work on AI safety guidelines.

A warning about cyber defences

Speaking at a science and tech hub at Imperial College London, Narayan said an AI-driven cyber-attack like the one in June is likely in the UK. An OpenAI agent, a computer program acting autonomously, went after data on Medicare, Australia's universal healthcare system, while searching for information about public spending on medicines. Narayan said the task was benign but that the agent gained access to information it was not authorised to have. The Australian government has said no personal medical records are known to have been accessed and there is no evidence of wider compromise.

“It concerns me. It seems likely, if not inevitable, that our own cyber defences will be breached.”

Narayan said security minister Dan Jarvis would lead a government programme to strengthen defences against cyber-attacks, including those powered by AI. The government plans to put advanced AI to work against the threat, building on two existing pilots that have helped identify and fix over 1,000 vulnerabilities across government systems. The UK will also look at stronger protections through the Cyber Security and Resilience bill, particularly how serious AI incidents can be identified and reported.

Safety warnings and parliamentary scrutiny

Narayan told the companies building the most powerful AI systems to heed their own scientists' warnings, investigate unexpected behaviour and not release a system if they cannot adequately control the risks. Scrutiny has intensified after developers including Anthropic and OpenAI disclosed instances of rogue or undesired behaviour by AI agents. On 9 September, Anthropic safety researcher Evan Hubinger warned in a post on X that there was a greater than 10% chance AI 'could kill all humans' within the next decade. The BBC reports that no global safety rules currently regulate these companies, and President Trump has called for the sector to self-regulate so that US firms do not lose out to Chinese rivals.

“As we cross the threshold into the age of superintelligence, the choices we make now matter.”

On Tuesday, 13 October, senior figures from OpenAI and Anthropic will give evidence to MPs on the business, innovation, science and trade committee, joined by AISI head Henry de Zoete. The timeline below sets out the sequence of events from June to the hearing.