AI Security Institute

The AI Security Institute (AISI) is a UK government research body established on November 2, 2023, under the Department for Science, Innovation and Technology (DSIT). It emerged from the world's first major AI Safety Summit at Bletchley Park, originally named the AI Safety Institute before rebranding. Its mission is to provide governments with a scientific understanding of AI's most serious risks and to develop practical safeguards. The New York Times has described it as a mix of weapons inspectors, epidemiologists, and code breakers, and it is one of the largest and best-funded AI safety research organisations globally.

The institute operates like a startup within government, blending official authority with private-sector agility. Adam Beaumont, former chief AI officer at GCHQ, serves as interim director, with Jade Leung, the Prime Minister's AI advisor, as Chief Technology Officer. AISI has signed agreements with Anthropic, Google, and OpenAI to test their frontier models before release. It offers an open-source platform called Inspect for standardised safety testing. In one notable instance, AISI identified serious vulnerabilities that could enable the development of biological weapons, which were fixed before the model launched.

AISI's research spans diverse AI applications. A study found that LLMs post-trained for political persuasiveness became less accurate and up to 51% more persuasive on political issues. Another report noted that nearly 10% of UK citizens use chatbots for emotional support weekly, and that AI systems now outperform PhD-level researchers on scientific knowledge tests. The institute maintains an active presence on X under @AISecurityInst and publishes incident reports and research findings publicly. [edit]