Senior Labour figures said at the party conference that AI safety is becoming a major concern, citing risks from extremist groups, cyberattacks and concentrated super-intelligent systems. One cabinet minister said the danger was not “killer robots” but powerful AI falling into the hands of bad actors. Another said the government was “falling behind the public” on AI safety and needed to catch up quickly.

Ed Miliband warned in his conference speech that technology company profits must not come at the expense of children or humanity. During the following seven days, AI became a central Westminster issue: AI minister Kanishka Narayan said rogue agents hacking the British state was “likely, if not inevitable,” and the government was seeking stronger cybersecurity legislation. AI was also discussed at cabinet on Tuesday. MI5 chief Ken McCallum said AI was “supercharging” the contest between cyber attackers and defenders, and warned that extremist groups could use it to unleash bioweapons.

Liam Byrne, who chairs Parliament’s business, innovation, science and trade committee, said the change from Keir Starmer’s government to Andy Burnham’s had shifted the government’s approach to AI threats. His committee heard evidence this week from civil service security leaders and Anthropic, OpenAI, DeepMind and Meta, whose representatives differed on whether government testing of new models should be mandatory. Byrne urged ministers to lead on regulation and work with the United States to monitor models. Reports said Donald Trump’s administration had asked AI companies not to share their latest models with the UK’s AI Security Institute for testing, but Narayan said cooperation had been retained. Byrne said ministers faced a choice between maintaining access to models and putting safeguards in place, as companies also navigate the White House.

The UK’s upcoming G20 presidency was described as an opportunity to push for international action. Geoffrey Irving, former chief scientist at the AI Security Institute, said commercial models from OpenAI and Anthropic had shown hacking capabilities but were released with safety guardrails. Freely available open source models can be adapted and, Irving said, typically caught up with paid models in less than a year. He said they may already be capable of serious harm, while safeguards on them may be absent or easy to remove.

Parliamentary witnesses warned that increasingly capable AI could help design novel biological harms, though experts said designing dangerous pathogens and turning them into weapons are separate steps. King’s College London professor Elena Simperl said there was “no evidence” for existential risk claims and urged attention to evidenced, immediate risks such as hacking and signs of AI addiction. Defence sources also raised concern about the cancellation last year of an AI programme at the Defence Science and Technology Laboratory at Porton Down. The Ministry of Defence said AI should be integrated across its work and cited £400m for AI-focused capabilities over the next four years.

Some officials worry that safety concerns could slow government adoption of AI tools for productivity and public services, including medical care and reducing waiting lists. Former cabinet secretary Gus O’Donnell said ministers should weigh the risk of being left behind alongside the need for safeguards, and cautioned that mistakes could damage public trust and set adoption back.

Senior Labour figures said at the party conference that AI safety is becoming a major concern, citing risks from extremist groups, cyberattacks and concentrated super-intelligent systems. One cabinet minister said the danger was not “killer robots” but powerful AI falling into the hands of bad actors. Another said the government was “falling behind the public” on AI safety and needed to catch up quickly.

Ed Miliband warned in his conference speech that technology company profits must not come at the expense of children or humanity. During the following seven days, AI became a central Westminster issue: AI minister Kanishka Narayan said rogue agents hacking the British state was “likely, if not inevitable,” and the government was seeking stronger cybersecurity legislation. AI was also discussed at cabinet on Tuesday. MI5 chief Ken McCallum said AI was “supercharging” the contest between cyber attackers and defenders, and warned that extremist groups could use it to unleash bioweapons.

Liam Byrne, who chairs Parliament’s business, innovation, science and trade committee, said the change from Keir Starmer’s government to Andy Burnham’s had shifted the government’s approach to AI threats. His committee heard evidence this week from civil service security leaders and Anthropic, OpenAI, DeepMind and Meta, whose representatives differed on whether government testing of new models should be mandatory. Byrne urged ministers to lead on regulation and work with the United States to monitor models. Reports said Donald Trump’s administration had asked AI companies not to share their latest models with the UK’s AI Security Institute for testing, but Narayan said cooperation had been retained. Byrne said ministers faced a choice between maintaining access to models and putting safeguards in place, as companies also navigate the White House.

The UK’s upcoming G20 presidency was described as an opportunity to push for international action. Geoffrey Irving, former chief scientist at the AI Security Institute, said commercial models from OpenAI and Anthropic had shown hacking capabilities but were released with safety guardrails. Freely available open source models can be adapted and, Irving said, typically caught up with paid models in less than a year. He said they may already be capable of serious harm, while safeguards on them may be absent or easy to remove.

Parliamentary witnesses warned that increasingly capable AI could help design novel biological harms, though experts said designing dangerous pathogens and turning them into weapons are separate steps. King’s College London professor Elena Simperl said there was “no evidence” for existential risk claims and urged attention to evidenced, immediate risks such as hacking and signs of AI addiction. Defence sources also raised concern about the cancellation last year of an AI programme at the Defence Science and Technology Laboratory at Porton Down. The Ministry of Defence said AI should be integrated across its work and cited £400m for AI-focused capabilities over the next four years.

Some officials worry that safety concerns could slow government adoption of AI tools for productivity and public services, including medical care and reducing waiting lists. Former cabinet secretary Gus O’Donnell said ministers should weigh the risk of being left behind alongside the need for safeguards, and cautioned that mistakes could damage public trust and set adoption back.