UK AISI vs US AISI vs EU AI Office (2026): The Three Frontier AI Regulators

UK AISI was announced at the Bletchley Park AI Safety Summit in November 2023, formally established in early 2024 inside DSIT (then 'Frontier AI Taskforce,' formalized as 'AI Safety Institute' to reflect a longer-term institutional commitment). Public website https://www.aisi.gov.uk/. The Institute's mandate is research, evaluation, and standards on frontier AI risks.

**Lab access.** UK AISI's most-cited differentiator is the depth of pre-deployment access it has secured from frontier labs. Per public lab blog posts and UK government statements, Anthropic, OpenAI, Google DeepMind, and other major frontier providers have granted UK AISI pre-deployment evaluation access to flagship models — including GPT-5, Claude Opus 4 and 4.7, and Gemini 2.5 Pro. The access is via voluntary agreements, not statutory powers.

**Research output.** UK AISI publishes methodology notes, evaluation findings, and research papers at https://www.aisi.gov.uk/work/our-publications. Topics covered include: bio capability evaluations, cybersecurity capability evaluations, model agentic and autonomy evaluations, jailbreak robustness, and safeguards research. The Institute also runs internal teams on agent evaluation, safeguards, criminal misuse, societal impacts, and cybersecurity.

**International coordination.** UK AISI co-leads the International AISI Network with US AISI and is a founding signatory. The Network coordinates joint evaluations, shares methodology, and produces joint communiqués (e.g. the AI Safety Summit follow-ups in Seoul 2024 and Paris 2025).

**Authority.** UK AISI does not have direct enforcement powers — it cannot levy fines on frontier providers, block deployments, or compel access. Its leverage comes from (a) the voluntary access agreements, (b) its role advising UK government on AI policy (including the AI Action Plan and any future UK AI legislation), and (c) the reputational weight of its findings. The UK government has signaled intent to put AI safety on a statutory footing but as of June 2026 no UK AI law equivalent to the EU AI Act has been enacted.

US AISI was announced in November 2023, formally established by NIST in early 2024 under the Department of Commerce. Public website https://aisi.nist.gov/. The Institute's mandate is to advance the science of AI safety, develop technical standards and tools, support evaluations, and convene the **AI Safety Institute Consortium** — a public-private partnership with 200+ member organizations including major AI providers, enterprise customers, academic institutions, and civil-society groups.

**Survived the 2025 EO rescission.** The October 2023 Biden Executive Order on Safe, Secure, and Trustworthy AI was substantially rescinded in January 2025. Some provisions (the 10^26 FLOPs reporting requirement, the dual-use foundation-model reporting) were withdrawn. Importantly, US AISI was established under separate NIST authority and persisted through the transition — the Institute continued operating, the AISI Consortium continued working groups, joint evaluations with UK AISI continued, and the International AISI Network coordination continued.

**Lab access.** Per public lab blog posts and US government announcements, OpenAI and Anthropic signed pre-deployment access agreements with US AISI in 2024, with subsequent evaluations of flagship models. Access is via voluntary agreements, similar to UK AISI.

**AISI Consortium.** The Consortium (https://aisi.nist.gov/aisic) is the most-cited US AISI mechanism for industry engagement. Working groups address: model testing and risk evaluation, safeguards, model transparency, synthetic content, capacity-building, and governance. Member organizations include the major frontier-AI providers, large enterprise users (Microsoft, IBM, AWS, etc.), academic institutions, and civil-society organizations. Output includes guidance documents and joint methodology that informs the NIST AI RMF generative-AI profile.

**Authority.** US AISI does not levy fines or block deployments — it operates as a standards-setting and evaluation body within NIST. Its leverage comes from (a) NIST's role in setting voluntary US technical standards (NIST AI RMF is the most-referenced US AI governance framework), (b) federal-agency procurement that increasingly aligns with NIST standards, (c) the AISI Consortium as a venue for industry-government coordination.

The EU AI Office was announced in January 2024 alongside the political agreement on the EU AI Act and became operational in February 2024 inside DG CNECT in the European Commission. Public website https://digital-strategy.ec.europa.eu/en/policies/ai-office. Its mandate is to operationalize the EU AI Act, with specific responsibility for General-Purpose AI provisions.

**Direct enforcement authority.** Unlike UK AISI and US AISI, the EU AI Office has statutory enforcement authority under the AI Act (Articles 88 and 89). It can investigate GPAI providers, request information and documentation, conduct evaluations, and recommend penalties up to €15M or 3% of global annual turnover for GPAI obligations violations. Penalties are formally imposed by the Commission. The Office also coordinates with national market surveillance authorities for high-risk AI system enforcement (high-risk enforcement is primarily at the member-state level).

**GPAI Code of Practice.** The Office facilitated the development of the GPAI Code of Practice — a multi-stakeholder document setting out concrete compliance pathways for GPAI obligations under Articles 53-55 of the AI Act. The Code was finalized in 2025 with input from frontier providers, civil society, academia, and downstream users. Signatories voluntarily commit to specific compliance pathways; the Code provides a presumption of conformity for compliant providers.

**Training-data summary template.** Article 53(1)(d) of the AI Act requires GPAI providers to publish 'a sufficiently detailed summary about the content used for training of the general-purpose AI model.' The AI Office finalized the template for this summary in 2025; every frontier provider serving the EU has been publishing per-model training-data summaries using this template.

**European Artificial Intelligence Board (EAIB).** The Office serves as secretariat for the EAIB — a body composed of representatives of each EU member state plus the European Data Protection Supervisor. The EAIB advises the Commission on AI Act implementation, coordinates national enforcement, and adopts guidance.

**AI Pact.** The Office launched the AI Pact in 2024 — a voluntary commitment program inviting providers to start AI Act compliance ahead of statutory deadlines. The Pact has dozens of signatories including major frontier providers and enterprise users.

The **International AISI Network** is the coordination layer that knits the three institutions plus other national bodies together. Founding signatories include UK AISI, US AISI, EU AI Office, plus AI safety bodies in Japan, Singapore, Canada, France, Australia, the Republic of Korea, and Kenya (per the Seoul Declaration follow-ups). The Network coordinates joint evaluations, shares methodology, and produces joint communiqués at the AI Safety Summits (Bletchley 2023, Seoul 2024, Paris 2025).

**Joint evaluations.** UK AISI and US AISI have publicly conducted joint pre-deployment evaluations of frontier models — e.g. the joint evaluations of Claude Opus 4 / 4.7 and GPT-5. Joint evals deduplicate effort and produce coordinated findings; per public communiqués, the Network has been moving toward standardized evaluation methodology.

**Coordination tensions.** The three institutions have different legal postures, different funding, different cultural orientations (research-led at UK AISI, standards-led at US AISI, enforcement-led at EU AI Office). Coordination requires reconciling these differences. Public outputs to date suggest a working model: joint methodology where possible, parallel work where appropriate, recognition of each institution's distinct mandate.

**Bilateral lab agreements.** Most frontier-lab pre-deployment access agreements are bilateral with each institute, not with the Network as a whole. So OpenAI's agreements with UK AISI and US AISI are separate documents; Anthropic's are separate; etc. Network coordination helps deduplicate evaluation activity but does not centralize access.

**Outside-Network bodies.** Several jurisdictions have AI safety + governance bodies that operate outside the formal Network — China's various AI standards bodies, India's IndiaAI mission, ASEAN frameworks. International coordination beyond the Network is a topic of ongoing diplomacy.

**UK AISI.** Publications at https://www.aisi.gov.uk/work/our-publications. Notable outputs: pre-deployment evaluation reports (with redactions for sensitive details), methodology notes on bio, cyber, agentic, and safeguards evaluations, AISI research papers (typically posted to arXiv with AISI affiliation), and joint outputs with US AISI and METR.

**US AISI.** Publications at https://aisi.nist.gov/publications. The AISI Consortium working groups publish reports, evaluation methodology, and standards drafts. Notable outputs: NIST AI RMF generative-AI profile (Jan 2024), AISI methodology guidance, joint evaluation reports with UK AISI.

**EU AI Office.** Publications and guidance at https://digital-strategy.ec.europa.eu/en/policies/ai-office and https://artificialintelligenceact.eu/. Notable outputs: GPAI Code of Practice (2025), training-data summary template (2025), AI Act implementation guidance, AI Pact materials, EAIB opinions and recommendations.

**International AISI Network.** Communiqués and joint statements published at AI Safety Summit follow-up venues. Network membership and shared methodology summaries available via national-institute pages.

**Reading order we recommend.** Start with the EU AI Office's GPAI Code of Practice and training-data summary template — they are the most concrete operational artifacts. Then read the UK AISI methodology notes for the cleanest published technical descriptions of frontier-model evaluations. Then read the US AISI Consortium working-group outputs for the most US-procurement-relevant guidance. Cross-reference against the relevant lab's Capability Reports / system cards (Anthropic, OpenAI, etc.).

**The access spectrum.** Pre-deployment access ranges from 'API access to a near-final model checkpoint, days before public release' to 'deep model-internal access including activations, gradients, weights, and engineering team interviews.' Most lab–institute agreements as of 2026 sit closer to the first end of the spectrum for frontier models, with deeper access for specific research collaborations.

**What evaluators do with access.** Capability evaluations (bio, cyber, agentic, persuasion, autonomy) using standardized eval suites. Safeguards testing (jailbreak attempts, prompt injection, instruction-hierarchy bypass). Red-teaming against specific misuse vectors. Reporting findings to the lab pre-deployment and (with lab consent + redaction) to the public post-deployment.

**Notable public summaries.** Anthropic publishes Capability Reports + Safeguards Reports for ASL-3+ models that summarize what UK AISI / US AISI / METR / Apollo found. OpenAI's system cards summarize what the same external evaluators tested. Both labs typically redact specific failure modes that could be misused, while publishing methodology and headline findings.

**Tension.** External evaluators sometimes want to publish more than labs want to share; labs sometimes want to publish faster than evaluators are ready to sign off. Public reporting to date suggests a working norm: evals are time-boxed (weeks-to-months pre-deployment), findings are jointly published in lab artifacts with evaluator attribution, evaluator-only research goes to arXiv after embargo.

**The institutional ask.** All three institutes have publicly stated that statutory pre-deployment access (not voluntary) would strengthen their work. The EU AI Office has the most direct path to this — the AI Act gives it information-gathering powers over GPAI providers. UK AISI and US AISI would require new legislation in their respective jurisdictions; both have signaled this as a longer-term goal.

**OpenAI.** Pre-deployment access agreements with UK AISI and US AISI (per OpenAI's safety blog at openai.com/safety). Engagement with EU AI Office on GPAI Code of Practice and training-data summaries. Member of the AISI Consortium. Signatory to the White House Voluntary AI Commitments.

**Anthropic.** Pre-deployment access agreements with UK AISI and US AISI (per Anthropic blog posts at anthropic.com/news). Engagement with EU AI Office on GPAI Code of Practice and training-data summaries. Member of the AISI Consortium. Signatory to Voluntary AI Commitments. Anthropic's RSP commits to publishing Capability + Safeguards Reports including third-party evaluator findings.

**Google DeepMind.** Engagement with UK AISI (per UK government communiqués and DeepMind blog posts). Engagement with EU AI Office. Member of the AISI Consortium. Signatory to Voluntary AI Commitments. DeepMind publishes the Frontier Safety Framework that intersects with these institutional surfaces.

**Microsoft.** Engagement with all three through enterprise + OpenAI integration. Member of the AISI Consortium. Engages on EU AI Act compliance via Azure / Microsoft 365 Copilot.

**Meta.** Engagement with EU AI Office on GPAI Code of Practice (Llama is a GPAI model). Member of the AISI Consortium. Signatory to White House Voluntary Commitments.

**Other providers.** Cohere, Mistral, Stability AI, IBM, Salesforce, and others engage with one or more of the three institutes via the AISI Consortium, the AI Pact, or direct provider-to-institute agreements.

**If you procure AI services.** The three institutes' published artifacts are the canonical reference for evaluation methodology. When you ask a vendor 'what evaluations have been done on this model,' the answer should reference UK AISI / US AISI methodology, METR or Apollo evaluation summaries, and the vendor's own Capability Report / system card. Vendors that engage substantively with these institutes can show you the artifacts.

**If you do internal evaluations.** UK AISI's published methodology notes and the US AISI Consortium working-group outputs are the best public starting points. Our Build LLM Red-Team Suite 2026, Run Anthropic Evals Locally, and LLM Jailbreak Detection with Promptfoo walk through the practical builds.

**If you serve EU customers.** The EU AI Office's GPAI Code of Practice + training-data summary template are mandatory reading for your AI provider diligence. Ask your provider for their training-data summary; it must use the AI Office template. Cross-reference against their Capability Report / system card and against the EU AI Act high-risk + GPAI obligations.

**If you do US federal procurement.** The NIST AI RMF (now with the generative-AI profile from US AISI Consortium work) is the most-referenced US framework. Federal agencies increasingly require NIST AI RMF alignment in procurement; the OMB memos M-24-10 and M-24-18 on agency AI use cite NIST AI RMF.

**If you build safety tooling.** All three institutes publish methodology that becomes the reference standard. Building your tools to match their methodology means your output is interoperable with the institutional reference.