Initiative overview
The AI Security Institute was established following the Bletchley Park AI Safety Summit in November 2023, building directly on the Foundation Model Taskforce which had operated from April 2023. It was the first government-backed AI safety body in the world and served as the model for equivalent institutions subsequently established in the United States, Japan, Canada, France, South Korea and Singapore under the International Network of AI Safety Institutes (INAIS). In January 2025, the incoming Labour government renamed the Institute the AI Security Institute, broadening its mandate to encompass AI-enabled cybersecurity threats, offensive AI capabilities, and adversarial uses of AI systems alongside the original focus on frontier model safety evaluation. Core activities include: conducting pre-deployment and post-deployment evaluations of frontier AI models from developers including Anthropic, OpenAI, Google DeepMind, Meta and others; publishing public-facing safety reports and evaluation methodologies; developing benchmarks and tooling for model assessment; contributing to international AI safety standards through INAIS, the OECD, ISO/IEC JTC 1/SC 42 and G7/G20 processes; and conducting research into systemic risks from advanced AI including misuse potential, deceptive alignment and emergent capabilities. The Institute operates the AI Safety Evaluations Platform (AISEP), an infrastructure enabling both UK government and invited international partners to run standardised model evaluations.
