---
id: ORG-SAFETY-001
title: "AI Safety Institutes and Research Organizations"
alternative_names:
  - "AI Safety Institutes"
  - "AI Security Institutes"
  - "Independent AI Evaluators"
category: "People & Organizations → Organizations"
volume: X
chapter: 56
entry_type: "Standard Article"
level: 3
status: "Draft"
last_reviewed: 2026-10-18
version: "0.1.0"
prerequisites:
  - "SAFE-ALIGN-001"
related_entries:
  - "ORG-LABS-001"
  - "ORG-STANDARDS-001"
  - "SAFE-XRISK-001"
  - "SAFE-EVAL-001"
  - "GOV-INTL-001"
  - "GOV-US-001"
---

# AI Safety Institutes and Research Organizations

## 1. Concise Definition
**AI safety institutes** are government-backed technical bodies that evaluate advanced AI systems and advise policymakers; **AI safety research organizations** are mostly non-governmental nonprofits that conduct or publish safety-relevant research and model evaluations. This article describes a selection of each, restricted to facts about their founding, stated mission and public outputs that are traceable to official pages, as of 2026-10-15.

## 2. Overview & Significance
Since 2023 a number of public bodies have been created to test frontier models, and a parallel set of independent organizations has specialized in evaluating whether AI systems can carry out dangerous or autonomous tasks. Their significance for this encyclopedia is evidential: when a developer reports on its own system (see [ORG-LABS-001]), these bodies are among the few sources able to provide a second assessment. The selection here is illustrative, not a census. Absence from this article does not imply absence of importance, and the article makes no claim about the relative influence of any organization.

## 3. Historical Development
**United Kingdom.** The UK government's overview document *Introducing the AI Safety Institute* (updated 17 January 2024) presents the Institute as "the next step" after the Frontier AI Taskforce, described there as "the first state body dedicated to the safety of advanced AI", and after hosting an AI Safety Summit (Department for Science, Innovation and Technology, 2024). On 14 February 2025 the UK government announced that the AI Safety Institute had become the "UK AI Security Institute", with a new criminal-misuse team and a stated emphasis on national security and crime (Department for Science, Innovation and Technology, 2025). The Institute describes itself as "a research organisation within the UK government's Department for Science, Innovation and Technology" (AISI, n.d.).

**United States and international network.** On 20 November 2024 the U.S. Departments of Commerce and State co-hosted the inaugural convening in San Francisco of the International Network of AI Safety Institutes. The accompanying fact sheet names the United States as inaugural chair and lists initial members as Australia, Canada, the European Union, France, Japan, Kenya, the Republic of Korea, Singapore, the United Kingdom and the United States (NIST, 2024). A NIST news item of 25 September 2025 names the U.S. body the Center for AI Standards and Innovation (CAISI) (NIST, 2025). As read on 2026-10-18, NIST's own page for the body and a NIST news item dated 16 October 2026 call it the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) (NIST, 2026b; NIST, n.d.). NIST also renamed the former AI Safety Institute Consortium the NIST AI Consortium on 29 May 2026 (NIST, 2026a). [VERIFICATION NEEDED: the June 2025 Commerce statement that created CAISI could not be retrieved (commerce.gov blocks automated access), and no NIST or Commerce announcement of the later CAISSI name was found, so the exact dates and basis of both changes are unconfirmed.]

**Independent organizations.** The Center for AI Safety published a one-sentence statement: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war", signed by a list of researchers and public figures (Center for AI Safety, n.d.). [VERIFICATION NEEDED: publication date of the statement, which the page read does not display; it was released in 2023 according to general recollection, not a source.] Apollo Research's own timeline places its first activities in 2023, including presenting at the UK AI Safety Summit and sharing recommendations with the Frontier AI Taskforce (Apollo Research, n.d.). Founding dates of METR and Redwood Research were not confirmed.

## 4. Technical / Conceptual Explanation
### 4.1 What institutes do
The UK Institute states three lines of work: "testing leading AI systems before they are released publicly and collaborating with top AI companies to improve their safety and security"; "informing policymakers across the UK and allied governments about emerging capabilities and risks"; and "advancing research into solutions with in-house research and >£15 million in grant funding" (AISI, n.d.). The figure is the Institute's own statement on a page read on 2026-10-15 and may change.

The International Network's fact sheet identified three priority topics for the inaugural meeting: managing risks from synthetic content, testing foundation models and conducting risk assessments for advanced AI systems (NIST, 2024).

### 4.2 Evaluation as a method
Safety institutes and evaluators typically run structured tests of a model's abilities (for example cyber capabilities or autonomous task completion) and report results to developers or the public. The method is described in [SAFE-EVAL-001]. A limitation common to all such tests is that an absence of demonstrated capability in a test does not establish absence of the capability.

### 4.3 Examples of published government evaluations
NIST's CAISI page, read 2026-10-15, lists several evaluations of open-weight models from developers based in the People's Republic of China, among them an evaluation of DeepSeek V4 Pro dated 1 May 2026, an assessment of Z.ai's GLM-5.2 dated 17 July 2026, a joint UK AISI and CAISI preliminary assessment of the cyber capabilities of Kimi K3 dated 23 July 2026, and an assessment of GLM-5.3 dated 17 September 2026 (NIST, n.d.). Only the listing was read; the content of those evaluations is not summarized here because the underlying reports were not read. [VERIFICATION NEEDED: findings of the cited evaluations]

## 5. Types, Components & Classifications
| Type | Description | Examples with sources above |
| :--- | :--- | :--- |
| Government institute | Technical body inside a national government | UK AI Security Institute; CAISI at NIST |
| Multilateral network | Coordinating forum of national institutes | International Network of AI Safety Institutes (launched 2024) |
| Independent evaluator | Nonprofit measuring model capabilities and risks | METR; Apollo Research |
| Safety research nonprofit | Conducts or funds research and advocacy | Center for AI Safety; Redwood Research |

METR describes itself as evaluating "frontier AI models to inform the public about their risks and capabilities", with most research assessing "the extent to which an AI system can autonomously carry out substantial tasks" (METR, n.d.). Redwood Research describes itself as "a nonprofit AI safety and security research organization" focused on risks that arise "if these powerful AI systems purposefully act against the interests of their developers and human institutions broadly" (Redwood Research, n.d.). Apollo Research states its goal as "to secure frontier AI systems across development, deployment, and governance" (Apollo Research, n.d.). These are self-descriptions.

## 6. Applications & Concrete Examples
**International reporting.** The *International AI Safety Report* (Bengio et al., 2025), authored by 96 listed contributors and chaired by Yoshua Bengio, is an example of a consolidated multi-author assessment. The report is a primary source for what its authors conclude; this article does not summarize its findings. [VERIFICATION NEEDED: report mandate, sponsoring governments and any later editions were not read]

**Pre-deployment testing.** The UK Institute's description of testing models "before they are released publicly" (AISI, n.d.) is the clearest official statement of pre-deployment evaluation in the sources read. The terms of access, and whether any developer has withheld or delayed a release as a result, are not established by these sources.

## 7. Advantages, Limitations & Unresolved Questions
**Advantages.** Public bodies can combine access to developers with the ability to inform policy; independent nonprofits can publish evaluations without a commercial relationship to a developer, although several describe collaboration with developers.

**Limitations.**
- Mandates, names and structures change quickly: the UK body was renamed in February 2025, about a year after the January 2024 overview document, and the U.S. body now carries a different name (see section 3).
- Self-descriptions and funding statements are developer-independent but not independently audited here.
- Evaluations cover a small fraction of released models, and the public record of them is incomplete.

**Unresolved questions.** (1) Whether voluntary pre-deployment testing reliably affects deployment decisions. (2) How comparable evaluations are across organizations. (3) How the shift in some institutes' emphasis from "safety" to "security" (the UK renaming is the documented example) changes the topics covered. These are open questions rather than findings; the sources read do not answer them.

## 8. Current State of Knowledge
As of 2026-10-15 the following are directly supported by official pages: the existence and self-described mission of the UK AI Security Institute; the November 2024 launch of the International Network with ten initial members; the existence of NIST's CAISI (shown on NIST's site as CAISSI on 2026-10-18) with a news list of evaluations of models from named developers; and the self-descriptions of METR, Redwood Research and Apollo Research. Independent academic assessment of the effectiveness of these bodies is not cited in this draft. See [GOV-INTL-001] for the governance context and [GOV-US-001] for U.S. policy.

## 9. Future Trajectory & Strategic Outlook
No registered forecasts. Observed trends: institutes expanding from model-testing toward security and misuse topics (UK renaming, 2025); joint testing across national bodies (the UK AISI and CAISI joint assessment listed by NIST); and continued growth of independent evaluators. Whether the International Network continues in its 2024 form is not established by the sources read; its current name and membership should be checked at the source. [VERIFICATION NEEDED: current status of the International Network as of 2026-10-15]

## 10. Related Entries & Knowledge Graph
- **Foundational Concepts**:
  - [[SAFE-ALIGN-001] AI Alignment](../vol08_safety_governance/SAFE-ALIGN-001_ai_alignment.md): The technical problem many of these organizations study.
  - [[SAFE-XRISK-001] Existential Risk from AI](../vol08_safety_governance/SAFE-XRISK-001_existential_risk.md): The risk framing used in the Center for AI Safety statement.
- **Related Methods & Technologies**:
  - [[SAFE-EVAL-001] Dangerous Capability and Alignment Evaluations](../vol08_safety_governance/SAFE-EVAL-001_dangerous_capability_and_alignment_evaluations.md): How evaluations are designed and their limits.
- **Institutions and Policy**:
  - [[ORG-LABS-001] Frontier AI Laboratories](./ORG-LABS-001_frontier_ai_laboratories.md): The developers whose systems are evaluated.
  - [[ORG-STANDARDS-001] Standards Bodies (ISO/IEC, NIST)](./ORG-STANDARDS-001_standards_bodies_iso_iec_nist.md): NIST's standards role alongside CAISI.
  - [[GOV-INTL-001] International AI Governance and Compute Governance](../vol08_safety_governance/GOV-INTL-001_international_ai_governance_and_compute_governance.md): Multilateral context.
  - [[GOV-US-001] United States AI Policy](../vol08_safety_governance/GOV-US-001_united_states_ai_policy.md): U.S. policy context.

## 11. References & Foundational Literature
1. AI Security Institute (AISI). (n.d.). *About*. https://www.aisi.gov.uk/about. [Self-described mission, three lines of work and grant funding statement; page read 2026-10-15.]
2. Apollo Research. (n.d.). *About Apollo Research*. https://www.apolloresearch.ai/about. [Stated goal and 2023 timeline; page read 2026-10-15.]
3. Bengio, Y., Mindermann, S., Privitera, D., et al. (2025). *International AI Safety Report*. arXiv preprint. arXiv:2501.17805. https://doi.org/10.48550/arXiv.2501.17805. [Multi-author report with 96 listed authors (per the DataCite record); cited only as an example of a consolidated assessment.]
4. Center for AI Safety. (n.d.). *Statement on AI Risk*. https://safe.ai/work/statement-on-ai-risk. [Text of the one-sentence statement.]
5. Department for Science, Innovation and Technology. (2024, updated January 17). *Introducing the AI Safety Institute*. GOV.UK. https://www.gov.uk/government/publications/ai-safety-institute-overview/introducing-the-ai-safety-institute. [Institute framed as the next step after the Frontier AI Taskforce.]
6. Department for Science, Innovation and Technology, AI Security Institute, & Kyle, P. (2025, February 14). *Tackling AI security risks to unleash growth and deliver Plan for Change*. GOV.UK. https://www.gov.uk/government/news/tackling-ai-security-risks-to-unleash-growth-and-deliver-plan-for-change. [Announcement that the UK AI Safety Institute became the UK AI Security Institute.]
7. METR. (n.d.). *About METR*. https://metr.org/about. [Self-description as an evaluator of frontier AI models; page read 2026-10-15.]
8. National Institute of Standards and Technology (NIST). (2024, November 20). *Fact Sheet: U.S. Department of Commerce & U.S. Department of State Launch the International Network of AI Safety Institutes at Inaugural Convening in San Francisco*. https://www.nist.gov/news-events/news/2024/11/fact-sheet-us-department-commerce-us-department-state-launch-international. [Launch date, inaugural chair, initial members and priority topics.]
9. National Institute of Standards and Technology (NIST). (2025, September 25). *CAISI Works with OpenAI and Anthropic to Promote Secure AI Innovation*. https://www.nist.gov/news-events/news/2025/09/caisi-works-openai-and-anthropic-promote-secure-ai-innovation. [Uses the name Center for AI Standards and Innovation (CAISI); page read 2026-10-18.]
10. National Institute of Standards and Technology (NIST). (2026a, May 29). *NIST Expands AI Consortium's Scope, Calls for New Members*. https://www.nist.gov/news-events/news/2026/05/nist-expands-ai-consortiums-scope-calls-new-members. [Renaming of the AI Safety Institute Consortium to the NIST AI Consortium; page read 2026-10-18.]
11. National Institute of Standards and Technology (NIST). (2026b, October 16). *Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation*. https://www.nist.gov/news-events/news/2026/10/key-practices-incident-readiness-and-response-super-intelligence-training. [Refers to "NIST's Center for Advancing Innovation and Standards for Super Intelligence (NIST CAISSI)"; page read 2026-10-18.]
12. National Institute of Standards and Technology (NIST). (n.d.). *Center for Advancing Innovation and Standards for Super Intelligence (CAISSI)*. https://www.nist.gov/caissi. [Page title and About text use the CAISSI name; news list of evaluations dated 2025-2026; the same body was listed as CAISI when the page was first read on 2026-10-15; page read again 2026-10-18.]
13. Redwood Research. (n.d.). *Home*. https://www.redwoodresearch.org/. [Self-description as a nonprofit AI safety and security research organization; page read 2026-10-15.]

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| UK Institute presented as successor step to the Frontier AI Taskforce | Established (government document) | Department for Science, Innovation and Technology (2024) | High |
| UK AI Safety Institute renamed AI Security Institute, 2025-02-14 | Established (government announcement) | Department for Science, Innovation and Technology et al. (2025) | High |
| International Network launched 2024-11-20 with ten initial members | Established (official fact sheet) | NIST (2024) | High |
| CAISI (named CAISSI on NIST's site, 2026-10-18) existence, evaluation listings and current name | Established (official pages, current) | NIST (2025); NIST (2026b); NIST (n.d.) | High for the listing; findings unread |
| Description of institute activities (testing, advising, grants) | Well Supported (self-description) | AISI (n.d.) | Moderate; outcomes not independently assessed |
| Independent evaluators' missions | Well Supported (self-description) | METR (n.d.); Redwood Research (n.d.); Apollo Research (n.d.) | Moderate |
| Effectiveness of pre-deployment testing | Contested / Not established | None cited | Open question |

## 13. Revision History
- **v0.1.0** (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.