The AI Security Institute, formerly the AI Safety Institute, is a British government body designed to understand the capabilities and impacts of advanced AI. It grew out of the Bletchley Park AI safety summit convened by Rishi Sunak in 2023 and was modelled on the pandemic-era vaccine task force that helped Britain become the first country to approve a clinically tested covid-19 jab. Just as that task force was led by Dame Kate Bingham, an outsider investor, AISI was chaired by Ian Hogarth, a venture capitalist.
William Isaac, a principal scientist at Google DeepMind, has called AISI "the crown jewel of all of the safety institutes". It became a template for similar outfits in America, Japan, Singapore and elsewhere. Britain chairs a global network of AI safety institutes.
AISI tests what AI models can do in dangerous domains such as cyber-security, biology and chemistry. Drawing on military war-gaming, it deploys "red teams" which pose as adversaries to see whether real rogues could bypass safety controls. In 2025 OpenAI reported that AISI had uncovered more than a dozen vulnerabilities that could have enabled users to develop biological weapons; these were fixed before the model's launch. AISI has secured access agreements with Anthropic, Google and OpenAI to test their models before release. Because leading labs refuse to grant access to their models' "weights", AISI's teams mostly evaluate models by asking them questions and seeing how they reply.
The institute's open-source platform, Inspect, allows firms, academics and governments to run standardised safety tests against a common yardstick.
One AISI study found that conversations with an AI chatbot were around 50% more persuasive at shifting political opinions than reading a static, AI-generated message. Another found that a third of British adults had used AI for emotional needs.
AISI received £100m in upfront government funding and runs on roughly £66m a year. Its interim director, Adam Beaumont, comes from GCHQ, Britain's signals-intelligence agency. Its chief technical officer is Jade Leung, who also serves as the government's AI adviser.
AISI tested Anthropic's Mythos model and found it neck-and-neck with other models on relatively simple cyber-security tests, but noticeably ahead in a more advanced test that requires a model to complete dozens of steps before successfully taking over a target machine.
At America's urging after the 2025 AI summit in Paris, AISI removed references on its website to "societal impacts" and to risks such as inequality and harm to welfare. A minister announced the institute would be renamed, the "S" denoting "security" rather than "safety". The changes were largely cosmetic; AISI's core focus had always been on security.
Youth is such a wonderful thing. What a crime to waste it on children.