AI Security Institute
In short
The AI Security Institute (AISI) is a UK government research body that tests the most advanced AI models for serious risks, such as help with cyber attacks or weapons. It was called the AI Safety Institute until February 2025.
- Also called
- AISI, UK AISI, AI Safety Institute
- Topic
- Law, policy and the UK
- Level
- Beginner
AISI sits inside the Department for Science, Innovation and Technology. It describes itself as a research organisation, not a regulator, and says it has more than 100 technical staff and £66m of funding a year1.
It began life in November 2023 as the AI Safety Institute, an evolution of the government's Frontier AI Taskforce. Its first three jobs were testing advanced AI systems, doing basic safety research and sharing what it found. The founding paper said plainly that it "is not a regulator and will not determine government regulation"2.
On 14 February 2025 the Technology Secretary, Peter Kyle, renamed it the AI Security Institute in a speech at the Munich Security Conference. The new focus is serious risks with security implications: AI used to develop chemical and biological weapons, carry out cyber attacks, or enable fraud and child sexual abuse. The announcement said it "will not focus on bias or freedom of speech", and set up a criminal misuse team with the Home Office3.
Its main tool is testing models before and after release. AI companies give it early access to leading models1. It also publishes free evaluation software, including a framework called Inspect1.
AISI has no power to block a model. Companies take part voluntarily. As of October 2026 there is no government AI bill; the AI bills listed by Parliament are private members' bills, such as Lord Holmes's Artificial Intelligence (Regulation) Bill, which has not got past first reading5.
Example
In November 2024 the UK and US institutes jointly tested Anthropic's upgraded Claude 3.5 Sonnet before it went public, checking its biological and cyber capabilities, software and AI development skills, and how well its safeguards held. The findings went to Anthropic before release4.
Why it matters
It is the UK state's main window into what the most powerful AI models can do before the public gets them. Its findings feed government decisions on security and regulation.
Questions people ask
Why did the AI Safety Institute change its name?
In February 2025 the government recast it as the AI Security Institute to focus on serious risks with security implications, such as AI-assisted weapons, cyber attacks, fraud and child sexual abuse material. The government said it would not focus on bias or freedom of speech3.
Is the AI Security Institute a regulator?
No. It was set up as a research and testing body that informs government, and its founding paper says it will not determine regulation2. AI companies give it access to models voluntarily1.
Who runs the AI Security Institute?
It is part of the UK government, sitting within the Department for Science, Innovation and Technology1.
Related terms
Sources
- AI Security Institute, About
- DSIT (2023), Introducing the AI Safety Institute, policy paper presented to Parliament
- DSIT press release (14 February 2025), Tackling AI security risks to unleash growth and deliver Plan for Change
- US AISI and UK AISI (2024), Pre-deployment evaluation of Anthropic's upgraded Claude 3.5 Sonnet
- UK Parliament Bills, Artificial Intelligence (Regulation) Bill [HL], sponsor Lord Holmes of Richmond, first reading 4 March 2025