AI Safety Institute Landscape — Seed Dossier

Date: 2026-05-24 Type: Research Seed Status: Seed dossier for NotebookLM deep research

Executive Summary

1. AISIs are a novel governance model — the first global-scale institutional framework for AI safety, combining technical evaluation capability within government structures while maintaining independence from direct regulatory power.
2. 10+ institutes now exist — US (CAISI, formerly US AISI), UK (renamed AI Security Institute), EU AI Office, Canada (CAISI), Japan (J-AISI), South Korea (Korea AISI), Singapore (Digital Trust Centre), Australia (announced Nov 2025, operational early 2026), Kenya, and France.
3. The International Network of AI Safety Institutes (renamed "International Network for Advanced AI Measurement, Evaluation and Science" — INAMAES) was announced at the Seoul Summit (May 2024) and held its inaugural meeting in San Francisco (Nov 2024) with 9 countries + EU.
4. Major shift from "safety" to "security" — the UK rebranded to "AI Security Institute" in 2025, the US moved from AISI to CAISI under NIST with a focus on national security evaluation and adversarial assessment of Chinese AI systems.
5. The 2026 International AI Safety Report (chaired by Yoshua Bengio) was published with contributions from 30+ countries, providing the most comprehensive global assessment to date.
6. Key critiques persist: lack of regulatory teeth, voluntary nature of evaluations, potential industry capture, underfunding, and questions about whether the shift to "security" dilutes the original safety mandate.

State of Play

Timeline of Key Events

Date	Event
Apr 2023	UK establishes Foundational Model Taskforce (precursor to AISI)
Nov 2023	UK AI Safety Summit at Bletchley Park; Bletchley Declaration signed by 28 countries
Nov 2023	US AISI announced within NIST (Executive Order 14110)
Jan 2024	UK AISI formally launched
Feb 2024	Japan AISI (J-AISI) established
May 2024	AI Seoul Summit; Seoul Statement of Intent signed; International Network of AISIs announced
Nov 2024	Inaugural meeting of International Network in San Francisco (9 countries + EU)
Feb 2025	Paris AI Action Summit; first International AI Safety Report published
Jun 2025	US restructures AISI as CAISI (Center for AI Standards and Innovation) under Commerce Secretary Lutnick
2025	UK rebrands to "AI Security Institute" with new £15M funding programme
Nov 2025	Australia announces its AI Safety Institute
Jan 2026	South Korea AI Basic Act takes full effect; Korea AISI formally established
Mar 2026	J-AISI publishes National Status Report on AI Safety in Japan 2025
May 2026	International AI Safety Report 2026 published (30+ countries)

Individual Institutes

1. United States — Center for AI Standards and Innovation (CAISI)

• Parent: NIST, Department of Commerce
• Founded: Originally US AISI (Nov 2023), restructured as CAISI (Jun 2025)
• Mandate: Serve as industry's primary point of contact for testing and collaborative research; develop guidelines for AI security; lead unclassified evaluations of AI capabilities posing national security risks; evaluate US and adversary AI systems; assess security vulnerabilities and malign foreign influence from adversary AI systems
• Key activities:
• Evaluated DeepSeek V4 Pro (Apr 2026)
• Signed CRADA with OpenMined for secure AI evaluations (Mar 2026)
• Signed MOU with GSA for AI evaluation in federal procurement (Mar 2026)
• Published NIST AI 800-4 on monitoring deployed AI systems (Mar 2026)
• Focus on US vs PRC AI capability comparison
• Key personnel: Led by NIST; Commerce Secretary Howard Lutnick oversees
• Notable shift: From "safety" to "standards and innovation" — explicit mandate to "guard against burdensome and unnecessary regulation" and "ensure US dominance of international AI standards"

2. United Kingdom — AI Security Institute (formerly AI Safety Institute)

• Parent: Department for Science, Innovation and Technology (DSIT)
• Founded: Apr 2023 (as Foundational Model Taskforce), rebranded 2025
• Mandate: "Minimise surprise to the UK and humanity from rapid and unexpected advances in AI"
• Key activities:
• Published approach to evaluations (Feb 2024)
• Systemic AI safety fast grants programme
• £15M funding programme for risks from misaligned AI systems
• Served as Secretariat for International AI Safety Report 2026
• Developed "Inspect" evaluation framework (open-source AI model evaluation toolkit)
• Key personnel: Originally led by Ian Hogarth; now part of DSIT
• Notable shift: Renamed from "Safety" to "Security" in 2025, reflecting broader UK government pivot toward AI security and national security framing

3. European Union — European AI Office

• Parent: European Commission
• Founded: 2024
• Structure: 125+ staff across 6 units + 2 advisors:
• Excellence in AI and Robotics
• Regulation and Compliance
• AI Safety
• AI Innovation and Policy Coordination
• AI for Societal Good
• AI in Health and Life Science
• Lead Scientific Advisor
• International Affairs Advisor
• Mandate: Implement the AI Act (especially GPAI rules); evaluate and classify GPAI models; draw up codes of practice; enforce GPAI model rules; promote trustworthy AI ecosystem
• Key powers: Can conduct evaluations of GPAI models, request information from providers, apply sanctions
• Unique feature: Regulatory teeth through AI Act enforcement — unlike most other AISIs, can directly sanction model providers

4. Canada — Canadian AI Safety Institute (CAISI)

• Parent: Innovation, Science and Economic Development Canada (ISED), with National Research Council and CIFAR
• Founded: 2024 (founding member of International Network)
• Mandate: Advance science of AI safety; understand and act on risks of advanced AI systems; focus on synthetic content risks, dangerous systems, human oversight
• Approach: Dual research model — applied/investigator-led research through CIFAR + government-directed projects
• Key activities: Joint testing with international counterparts; bilateral MOU with Australia; founding member of International Network
• Research focus: Risks from synthetic content (impersonation, fraud)

5. Japan — Japan AI Safety Institute (J-AISI)

• Parent: Cabinet Office (cross-ministry coordination)
• Founded: Feb 2024
• Mandate: Study and promote evaluation methods and standards for AI safety; serve as hub for AI safety information in Japan; establish international consensus through collaboration
• Key activities:
• Published "Guide to Evaluation Perspectives on AI Safety" (Version 1.10, revised for multimodal)
• Published National Status Report on AI Safety in Japan 2025 (Mar 2026)
• Systematization of evaluation methods and guides
• Standards and conformity assessment development
• Deepening international cooperation
• Unique approach: Operates as part of a government agency, enabling direct policy influence; cross-ministry coordination through Cabinet Office
• Key insight (from report): "There has been no precedent for cross-national or global coordination efforts on AI safety, and countries are still searching for ways to create a structure that integrates policy and technology."

6. South Korea — Korea AI Safety Institute

• Legal basis: Article 12 of the AI Basic Act (passed Dec 2024, in full effect Jan 2026)
• Mandate: Secure AI safety; protect people's lives, physical well-being, property from AI risks
• Key activities: Bilateral AI evaluation testing with Singapore AISI; collaboration on joint evaluation frameworks
• Governance context: AI Basic Act also establishes National AI Strategy Committee (chaired by president) and AI Policy Center

7. Singapore — Digital Trust Centre / AI Safety Institute

• Key activities: Bilateral testing collaboration with Korea AISI; Concordia AI published "State of AI Safety in Singapore 2025"
• Approach: Emphasis on digital trust, evaluation tools, and international interoperability

8. Australia — Australian AI Safety Institute

• Parent: Department of Industry, Science and Resources
• Announced: 25 November 2025
• Operational: Early 2026
• Mandate: Monitor, test, and share information on emerging AI technologies, risks, and harms; help government keep pace with rapid AI developments; serve as central hub for coordinated government action
• Key activities:
• Join International Network (INAMAES)
• Work with National AI Centre (NAIC)
• Support Australia's commitments under international AI safety agreements
• Funding concerns: Critics note funding may be insufficient for the mandate

9. Kenya

• Status: Member of International Network (replaced Italy/Germany as final founding member)
• Significance: Represents Global South participation in AISI network; important for geographic diversity

10. France

• Status: Signatory to Seoul Statement; member of International Network
• Context: Hosted Paris AI Action Summit (Feb 2025); has own AI safety evaluation capabilities

The International Network (INAMAES)

• Full name: International Network for Advanced AI Measurement, Evaluation and Science (formerly International Network of AI Safety Institutes)
• Founded: May 2024 (AI Seoul Summit)
• Founding members: US, UK, EU, Japan, Canada, Singapore, South Korea, France, Kenya, Australia
• Inaugural meeting: November 2024, San Francisco
• Mission: "Accelerate the advancement of the science of AI safety" at a global level by promoting "complementarity and interoperability" between institutes
• Activities: Coordinating research, sharing resources, developing best practices, exchanging or codeveloping AI model evaluations
• Key deliverable: Joint contributions to the International AI Safety Report (2025, 2026 editions)

Evaluation Frameworks and Tools

1. Inspect (UK AISI) — open-source AI model evaluation toolkit for safety testing
2. Dioptra (NIST/CAISI) — AI model evaluation and red-teaming framework
3. NIST AI 800-4 — Post-deployment monitoring of AI systems
4. J-AISI Guide to Evaluation Perspectives on AI Safety — systematized evaluation methodology
5. AI Act codes of practice (EU AI Office) — state-of-the-art codes for GPAI providers
6. CRADA with OpenMined (CAISI) — secure/privacy-preserving AI evaluation methods

Key Critiques and Controversies

1. Shift from "Safety" to "Security": UK and US rebranding toward security/national security framing raises concerns that the original safety mandate (focused on societal harms, bias, transparency) is being diluted in favor of cybersecurity and geopolitical competition.

2. Lack of regulatory teeth: Most AISIs cannot compel compliance — they evaluate but cannot block deployment. The EU AI Office is a notable exception with enforcement powers under the AI Act.

3. Voluntary model access: Evaluations depend on voluntary pre-deployment access agreements with AI companies. No jurisdiction mandates this (though the EU AI Act changes this for GPAI models).

4. Industry capture risk: Close working relationships with AI developers (OpenAI, DeepMind, Anthropic) raise concerns about regulatory capture.

5. Underfunding: Several institutes (notably Australia) face criticism that resources don't match the ambition.

6. US "standards dominance" agenda: CAISI's explicit mandate to "ensure US dominance of international AI standards" and "guard against burdensome regulation" appears to conflict with the cooperative, science-first framing of the network.

7. Evaluation methodology immaturity: As J-AISI notes, "countries are still searching for ways to create a structure that integrates policy and technology." Current evaluation methods (red-teaming, benchmarks) are acknowledged to be "easy to manipulate."

8. Geopolitical tensions: The US-China AI competition (evident in CAISI's DeepSeek evaluations and adversary-focused mandate) complicates the neutral, scientific framing of AISIs.

9. Paris Summit pivot: The Paris AI Action Summit (Feb 2025) shifted focus from safety to innovation/action, raising questions about political commitment to the safety agenda.

Unverified / Contested

• Exact budgets for most institutes (publicly undisclosed or partially disclosed)
• Internal evaluation results are often classified or shared only with model developers
• Whether voluntary evaluation agreements with frontier AI companies are being honoured in practice
• The degree to which the US shift to CAISI represents genuine institutional change vs. rebranding

Source List

Government/Official

1. International AI Safety Report 2026 — https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026
2. NIST CAISI — https://www.nist.gov/caisi
3. Australia AI Safety Institute — https://www.industry.gov.au/news/australia-establish-new-institute-strengthen-ai-safety
4. Japan AISI National Status Report 2025 — https://aisi.go.jp/assets/pdf/j-aisi_report_2025_en.pdf
5. Canada CAISI — https://ised-isde.canada.ca/site/ised/en/canadian-artificial-intelligence-safety-institute
6. European AI Office — https://digital-strategy.ec.europa.eu/en/policies/ai-office
7. Seoul Declaration — https://www.industry.gov.au/publications/seoul-declaration-countries-attending-ai-seoul-summit-21-22-may-2024
8. International Network Mission Statement — https://ised-isde.canada.ca/site/ised/en/international-network-ai-safety-institutes-mission-statement

Think Tanks / Analysis

1. CSIS: AISI International Network Next Steps — https://www.csis.org/analysis/ai-safety-institute-international-network-next-steps-and-recommendations
2. All Tech Is Human: Global Landscape of AISIs — https://alltechishuman.org/all-tech-is-human-blog/the-global-landscape-of-ai-safety-institutes
3. NBR: South Korea AI Governance — https://www.nbr.org/publication/south-koreas-approach-to-ai-governance-mitigating-risks-and-ensuring-ai-safety/
4. Concordia AI: State of AI Safety in Singapore 2025 — https://concordia-ai.com/wp-content/uploads/2025/07/State-of-AI-Safety-in-Singapore-2025.pdf

Video

1. Inside The Second Int'l AI Safety Report — https://www.youtube.com/watch?v=2VlXhGottLw
2. Max Tegmark on AI Safety Index 2025 — https://www.youtube.com/watch?v=hGUUhxNn86M
3. Australia $30M AI safety crackdown — https://www.youtube.com/watch?v=iNN24KYPxuA