UK AI Safety Institute

The UK AI Safety Institute (UK AISI) — renamed in February 2025 to the AI Security Institute, continuing under the same AISI acronym — is the British government's frontier-artificial-intelligence safety research body, established in November 2023 at the Bletchley Park AI Safety Summit by then-Prime Minister Rishi Sunak. The institute operates as a directorate of the UK Department for Science, Innovation and Technology (DSIT) with research staff in London and dedicated arrangements with frontier-AI labs that provide pre-deployment evaluation access to advanced models. UK AISI was the world's first government-backed AI safety institute and has been characterized in industry coverage as the principal government AI safety research body globally, with active pre-deployment evaluation cooperation across OpenAI, Anthropic, Google DeepMind, Meta AI / FAIR, Mistral AI, and adjacent leading AI labs. The institute's open-source Inspect AI safety evaluation framework has become widely adopted across the broader AI safety research community, and the institute publishes selected pre-deployment evaluation reports under public-facing transparency commitments. As of April 2026, UK AISI is the leading national-government AI safety research body globally and the principal counterpart to the US AI Safety Institute (NIST) under the bilateral cooperation arrangements.

At a glance

• Founded: November 2023 by the UK Government at the Bletchley Park AI Safety Summit. Renamed in February 2025 to the AI Security Institute, continuing as AISI.
• Status: UK government research directorate within the Department for Science, Innovation and Technology (DSIT). Headquartered in London.
• Funding: UK government funding through DSIT. Initial commitment of approximately £100 million reported at the November 2023 launch, with expansions through 2024 to 2026. Specific fiscal-year allocations have not been comprehensively disclosed.
• Chief Scientist: Geoffrey Irving, Chief Scientist of UK AISI (since the 2023 founding period). PhD mathematics (Stanford); long-tenured AI safety researcher with prior senior research roles at Google DeepMind (research scientist on AI safety) and OpenAI (research scientist on debate and AI safety alignment).
• Other notable leadership: Ian Hogarth, founding Chair of the UK Frontier AI Taskforce (the precursor body that became AISI); transitioned from operational leadership to advisory engagement after the institute's establishment. Yarin Gal, Director of Research; concurrent Associate Professor at Oxford with a research-leadership role anchored on the Inspect framework. Senior research-and-engineering staff including Mary Phuong (formerly Google DeepMind), Marius Hobbhahn (Apollo Research; AISI advisory), and adjacent senior recruits from frontier-AI labs and academic AI safety research.
• Open weights: Limited. UK AISI's research outputs are predominantly evaluation reports, AI safety research papers, and the Inspect open-source evaluation framework rather than foundation-model releases.
• Flagship outputs: Inspect (open-source AI safety evaluation framework, released May 2024); pre-deployment evaluation reports published for selected frontier AI models including OpenAI o1 (December 2024 joint UK AISI / US AISI evaluation report) and Anthropic Claude variants; AI safety research papers across dangerous-capability evaluation, biosecurity, cybersecurity, and adjacent areas; policy contributions to the Bletchley AI Safety Summit (November 2023), the Seoul AI Safety Summit (May 2024), and the Paris AI Action Summit (February 2025).

Origins

The UK AI Safety Institute's lineage traces to the UK Frontier AI Taskforce, established in April 2023 under the leadership of Ian Hogarth, the entrepreneur and AI investor who had publicly characterized AI risk as substantial in his Financial Times essay "We Must Slow Down the Race to God-like AI" (April 2023). The Frontier AI Taskforce was given an initial £100 million commitment by then-Prime Minister Rishi Sunak, with the explicit mandate to build UK government capability for evaluating frontier AI capabilities and to inform UK and international AI safety policy.

The November 2023 Bletchley Park AI Safety Summit was the principal early UK government convening on AI safety, with international government leaders (US Vice President Kamala Harris, EU Commission President Ursula von der Leyen, UN Secretary-General António Guterres) and frontier-AI lab CEOs attending alongside Sunak. The summit produced the Bletchley Declaration on AI safety, signed by 28 governments including the US, China, and the EU. Concurrent with the summit, Sunak announced the formal establishment of the UK AI Safety Institute as the successor body to the Frontier AI Taskforce, with the broader institutional mandate of pre-deployment frontier model evaluation, AI safety research, and international AI safety cooperation.

Geoffrey Irving was appointed Chief Scientist at the founding period, with the senior research-leadership role providing both technical research direction and a credible academic-research profile that anchored UK AISI's positioning within the broader AI safety research community. Irving's prior roles at OpenAI and Google DeepMind provided the frontier-lab-research credibility, and his Stanford mathematics PhD anchored the academic-research credibility.

The 2024 to 2025 period saw UK AISI expand staffing and research output substantially. The May 2024 release of the Inspect AI safety evaluation framework was the institute's principal open-source contribution and has since become widely adopted across the AI safety research community as the standard tooling for AI capability evaluation. Pre-deployment evaluation cooperation with leading frontier-AI labs produced selected publicly released reports including the December 2024 joint evaluation of OpenAI o1 (released jointly with the US AI Safety Institute under the bilateral cooperation arrangements). The May 2024 Seoul AI Safety Summit and the February 2025 Paris AI Action Summit continued the international-cooperation programs that the institute had anchored from inception.

In February 2025, the institute was renamed to the AI Security Institute, reflecting the broader UK government strategic shift toward national-security framing of AI capabilities. The acronym AISI was retained, with industry coverage characterizing the rename as a signal of the post-2024 UK government recalibration on AI policy from the Bletchley-era safety framing toward a security-and-capability framing more compatible with both the new UK Labour government's economic priorities and the post-January 2025 Trump administration's US AI policy direction.

The 2025 to 2026 period has continued frontier AI evaluation, Inspect framework development, and international-cooperation activity through the bilateral relationship with the US AI Safety Institute and adjacent international AI safety bodies.

Mission and strategy

UK AISI's stated mission is to advance the science of AI safety, with emphasis on pre-deployment evaluation of frontier AI capabilities, dangerous-capability research, AI safety methodology development, and international AI safety cooperation. The strategy combines four threads. First, pre-deployment frontier AI model evaluation through agreements with leading AI labs, providing UK AISI early access to frontier model capabilities under specified-confidentiality arrangements. Second, AI safety research output across dangerous-capability evaluations (biosecurity, cybersecurity, autonomy, persuasion), foundational AI safety methodology, and adjacent areas. Third, the Inspect AI safety evaluation framework as open-source infrastructure that enables broader AI safety research-community evaluation work. Fourth, policy engagement with the UK government, the EU, the US, and the broader international AI safety community.

The competitive premise is that frontier AI safety evaluation requires specialized capability that academic AI safety research cannot fully provide and that frontier-lab internal safety teams have structural conflicts of interest that an independent government research body can address, and that the UK's combination of academic-AI-research depth, frontier-lab cooperation arrangements, and government-research credibility can produce AI safety research outputs that no other institution globally can match.

Models and products

• Inspect. Open-source AI safety evaluation framework released May 2024. The principal AI safety evaluation tooling globally; widely adopted across academic and industry AI safety research.
• Pre-deployment evaluation reports. Selected reports published publicly including the December 2024 joint UK AISI / US AISI evaluation of OpenAI o1.
• AI safety research papers. Output across dangerous-capability evaluation, biosecurity, cybersecurity, autonomy, persuasion, and adjacent AI safety research areas.
• Policy and regulatory contributions. Contributions to the Bletchley Declaration (November 2023), the Seoul AI Safety Summit (May 2024), the Paris AI Action Summit (February 2025), and ongoing UK and international AI safety policy.
• Bilateral cooperation with the US AI Safety Institute (NIST). Memorandum of Understanding signed April 2024.

Distribution channels include open-source software (Inspect framework), publicly released evaluation reports, AI safety research papers through ArXiv and academic venues, and policy submissions to UK and international AI safety bodies.

Benchmarks and standing

UK AISI's evaluation framework focuses on the rigor and adoption of its AI safety evaluation methodology (with Inspect as the principal artifact), the quality of pre-deployment evaluation reports, the recruiting competitiveness for senior AI safety research talent, and the institutional credibility within the broader AI safety research community.

Industry coverage has consistently characterized UK AISI as the leading national-government AI safety research body globally, with the November 2023 Bletchley Summit founding context, the Geoffrey Irving research-leadership credibility, the Inspect framework adoption, and the substantive frontier-lab cooperation as principal validating data points. The February 2025 rename to the AI Security Institute has been characterized in industry coverage as a signal of strategic recalibration rather than a fundamental change in mission, with continued frontier AI evaluation work through 2025 to 2026.

Leadership

As of April 2026, UK AISI's senior leadership includes:

• Geoffrey Irving, Chief Scientist.
• Yarin Gal, Director of Research. Concurrent Oxford Associate Professor.
• Senior research-and-engineering staff across the dangerous-capability evaluation, biosecurity, cybersecurity, autonomy, and Inspect framework teams.

Ian Hogarth, founding Chair of the UK Frontier AI Taskforce, transitioned from operational leadership to advisory engagement after the institute's establishment.

Funding and backers

UK government funding through the Department for Science, Innovation and Technology (DSIT). Initial commitment of approximately £100 million reported at the November 2023 launch, with expansions through 2024 to 2026.

Industry position

UK AISI occupies a structurally distinctive position as the leading national-government AI safety research body globally, with the multi-year operating history (since November 2023), the substantive frontier-lab cooperation arrangements, the Inspect framework adoption, and the senior research-leadership credibility. The institute has been characterized in industry coverage as the principal government AI safety institution worldwide.

The structural risks are two. First, the post-2025 rename to the AI Security Institute and the broader UK government policy recalibration introduce uncertainty about the institute's continued mission scope and funding trajectory. Second, the bilateral cooperation with the US AI Safety Institute depends on continued US government commitment that has been less stable through the post-January 2025 Trump administration transition than the parallel UK-government commitment.

Competitive landscape

• US AI Safety Institute (NIST). Bilateral cooperation partner under the April 2024 Memorandum of Understanding. Different governance (NIST as a US Department of Commerce standards body) and shifting US-policy context.
• EU AI Office. European Union's AI safety oversight body within the European Commission. Different regulatory mandate.
• Anthropic AI safety teams, Google DeepMind AGI safety, OpenAI Superalignment / Safety Systems. Frontier-AI-lab internal safety research organizations.
• Apollo Research, METR, Transluce, Timaeus, Conjecture, Goodfire. Independent AI safety research peer organizations.
• MIRI (Machine Intelligence Research Institute), Center for AI Safety, CHAI (Berkeley Center for Human-Compatible AI). Earlier-cohort AI safety research organizations.
• Alan Turing Institute. UK national academic AI research institute. Adjacent UK AI ecosystem peer.

Outlook

• Continued frontier AI evaluation and research output through 2026 to 2027.
• The renamed AI Security Institute trajectory under the post-2025 framing.
• Bilateral cooperation with the US AI Safety Institute under the evolving US-policy context.
• Continued senior research-talent recruitment and retention against competing offers from frontier-AI labs and academic positions.
• The institute's role in shaping UK and international AI safety regulation through 2026 and 2027 international AI safety summits.

Sources

• UK AI Safety Institute official site. Institute reference.
• UK Government: AI Safety Institute announcement. November 2023 launch announcement.
• Inspect AI evaluation framework. Open-source evaluation framework reference.
• Bletchley Declaration. November 2023 summit declaration.
• Geoffrey Irving LinkedIn. Chief Scientist reference.