evaluator; distance from a lab: 1
UK AI Security Institute
Scorecard
Largest state evaluation team; 30+ models tested; Inspect open-sourced; Frontier AI Trends Report.
Government, London, UK. Government institute; listed with government institutes. Confidence high. Domains: cyber, bio, jailbreak, autonomy, assurance.
No floor triggered 8/8 evidenced · Retrieved & confirmed Scores by preset: Lab procurement –; Regulator or auditor –; Public trust –; Equal weights –. Weakest evidenced dimension: publication rights 2/4. No evidenced dimension below 2 — not a pass. No floor triggered means no disqualifying conflict is on file, not a pass. Full evidence.
The weighted numbers are hidden until you ask for them: two readers with different priorities see different numbers, and neither is the site's.
What would move the score. Statutory footing with power to compel testing and publish per-model findings.
- Funding 3 to 4: score –
- Governance 3 to 4: score –
- Personnel 3 to 4: score –
Dissent, lower. Publication (R) at 2 could be 1. The public record shows per-model findings from more than thirty tested systems have never been published; the only public output is one aggregated Trends Report, and CSA describes the underlying work as confidential between the Institute and individual labs. Under R.4 a public body that publishes nothing per model scores 1, and one aggregate report in two years is close to nothing per model.
Dissent, higher. Publication (R) at 2 could be 3. The Trends Report published an adverse finding against every lab at once: a universal jailbreak in every system tested. The joint evaluations with CAISI produced public write-ups of ChatGPT Agent vulnerabilities and Constitutional Classifier bypasses. No lab-edited summary is documented, and the anchor-3 text (publishes, security-only redaction) is closer to that record than lab-edited summaries.
Each value is the tightest admissible cap or, with no cap, the highest admissible floor, under RULES.md. The binding signal is highlighted. A dash means no admissible signal sets a bound under the Retrieved & confirmed policy.
Funding 3/4
Anchor 3: Mostly philanthropic or public money; some lab-linked pooled funds.
evidence: tier 3 (self)
Capped at 3 by ukaisi.11 (F.9): AISI administers the Alignment Project, a GBP 27m grant pool that includes GBP 5.6m from OpenAI and money from Anthropic, Microsoft and AWS; core budget of about GBP 66m per year is public.; capped at 3 by ukaisi.12 (F.9): The Alignment Project took £5.6m from OpenAI. The £27m/£12m totals, the wider backer list, and OpenAI’s non-influence statement are per the cited sources; spans not yet captured.. Floors up to 3 from ukaisi.01, ukaisi.15 do not exceed the cap.
Curator: Anchor 3: public core budget, but the institute runs a lab-funded grant pool, which is lab money passing through it even if not into its evaluation budget.
- Taxpayer funded; no commercial relationship with labs. floors at 3 by F.9“internal technical research directorate within the Department for Science, Innovation and Technology”
AI Security Institute (renaming) (Regulations.ai, 2026-09) tier 4 press (aggregator)confirmed About the AI Security Institute (UK AI Security Institute, 2026) tier 3 self (ukaisi)confirmed2026-09-14 - AISI administers the Alignment Project, a GBP 27m grant pool that includes GBP 5.6m from OpenAI and money from Anthropic, Microsoft and AWS; core budget of about GBP 66m per year is public. caps at 3 by F.9binding“OpenAI, Microsoft, Amazon Web Services (AWS), Schmidt Sciences, Anthropic”
Alignment Project grants (UK AI Security Institute, 2025) tier 3 self (ukaisi)confirmed About the Alignment Project (UK AI Security Institute, 2026) tier 3 self (ukaisi)confirmed About the AI Security Institute (UK AI Security Institute, 2026) tier 3 self (ukaisi)confirmed2026-09-14 - The Alignment Project took £5.6m from OpenAI. The £27m/£12m totals, the wider backer list, and OpenAI’s non-influence statement are per the cited sources; spans not yet captured. caps at 3 by F.9binding“including £5.6m from OpenAI”
Funding 60 projects to advance AI alignment research (UK AI Security Institute, 2026-02) tier 3 self (ukaisi)confirmed About the Alignment Project (UK AI Security Institute, 2026) tier 3 self (ukaisi)confirmed Advancing independent research on AI alignment (OpenAI, 2026) tier 3 self (openai)confirmed2026-02 - Alignment Project grants are contracted and paid by each funder directly to the host organisation; DSIT/AISI's own grants are statutory grants under section 5 of the Science and Technology Act 1965 managed by an AISI research lead, and other funders issue their own terms, so OpenAI's £5.6m is not received into AISI's budget even though AISI runs the selection. floors at 3 by F.9“the Funding Agreement (GFA) will be between the funder(s) and the host organisation only”
The Alignment Project - Clarification Questions V2.0 (PDF) (The Alignment Project (UK AI Security Institute), 2025-08) tier 3 self (ukaisi)confirmed How to apply (The Alignment Project (UK AI Security Institute), 2025-07-30) tier 3 self (ukaisi)confirmed Example of DSIT-AISI Grant Funding Agreement (PDF) (Department for Science, Innovation and Technology / AI Security Institute, 2025-08) tier 3 self (ukaisi)confirmed2026-09-15 - Document request: Whether Alignment Project funds touch evaluation staff or infrastructure.
Governance 3/4
Anchor 3: Nonprofit or public body with a COI policy.
evidence: tier 1 (filing/index)
Floored at 3 by ukaisi.08 (G.2): Public body with civil-service conflict rules.. No admissible signal caps it.
Curator: Anchor 3: public body with civil-service conflict rules (ukaisi.08). No tier-1 source for an independent board or external review, so 4 is unearned under the tier rule (C10).
- Public body with civil-service conflict rules. floors at 3 by G.2binding“The Institute is an advisory technical research body rather than a formal regulator”
AI Security Institute (renaming) (Regulations.ai, 2026-09) tier 4 press (aggregator)confirmed The Civil Service Code (GOV.UK (Cabinet Office), 2015-03-16) tier 1 filingconfirmed2026-09-14
Personnel 3/4
Anchor 3: Recusal policy and disclosure of lab ties.
evidence: tier 1 (filing/index)
Floored at 3 by ukaisi.13 (P.5): AISI staff are civil servants in DSIT and are bound by the Civil Service Code (conflicts and gifts) and the Business Appointment Rules, a statutory post-employment rule that applies for one to two years after leaving Crown service.. No admissible signal caps it.
- Staff move between the institute and labs in both directions. informational (P.3: two-way hiring caps at 2, but "a statutory post-employment rule for the body's staff lifts the cap"; the Business Appointment Rules (new signal ukaisi.13) lift it. Absent ukaisi.13 this signal caps P at 2.)
UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed2026-09-14 - AISI staff are civil servants in DSIT and are bound by the Civil Service Code (conflicts and gifts) and the Business Appointment Rules, a statutory post-employment rule that applies for one to two years after leaving Crown service. floors at 3 by P.5binding“two years after leaving Crown service, if you are grade SCS1 or SCS2”
The Civil Service Code (GOV.UK (Cabinet Office), 2015-03-16) tier 1 filingconfirmed Business Appointment Rules for Crown Servants: guidance (GOV.UK (Cabinet Office), 2025) tier 1 filingconfirmed About the AI Security Institute (UK AI Security Institute, 2026) tier 3 self (ukaisi)confirmed2026-09-15
Access depth (lab-granted) 3/4
Anchor 3: Helpful-only or weights-level access, chain of thought, logs, on-site.
evidence: tier 4 (press)
mechanism: lab-controlled
Floored at 3 by ukaisi.06 (A.4): Joint testing with the US institute; agreements with OpenAI, Anthropic, Google, Microsoft, Cohere; 30+ models tested.. No admissible signal caps it.
- No statutory powers; access rests on MOUs any lab could decline to renew; Frontier AI Bill not introduced as of September 2026. informational (A.3: "the lab can decline to renew" is recorded as a mechanism (lab-controlled), not as a cap.)“the bill had not yet been introduced to Parliament”
UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed AI Security Institute (renaming) (Regulations.ai, 2026-09) tier 4 press (aggregator)confirmed2026-09-14 - Joint testing with the US institute; agreements with OpenAI, Anthropic, Google, Microsoft, Cohere; 30+ models tested. floors at 3 by A.4binding“running pre-deployment and post-deployment evaluations of the most capable models released by major AI developers”
UK AI Security Institute (AI Wiki, 2026-05-07) tier 4 press (aggregator)confirmed UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed2026-09-14
Scope control 3/4
Anchor 3: Evaluator sets scope and can add questions.
evidence: tier 4 (press)
Capped at 3 by ukaisi.14 (S.4): No statutory power to define scope or compel testing; the institute operates through voluntary testing agreements backed by bilateral memoranda of understanding, and the Frontier AI Bill had not been introduced as of September 2026.. Floors up to 3 from ukaisi.09 do not exceed the cap.
- Sets its own evaluation agenda within MOU terms. floors at 3 by S.4“AISI’s voluntary testing arrangements with frontier developers remain the primary mechanism”
UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed2026-09-14 - No statutory power to define scope or compel testing; the institute operates through voluntary testing agreements backed by bilateral memoranda of understanding, and the Frontier AI Bill had not been introduced as of September 2026. caps at 3 by S.4binding“The institute cannot fine, audit, or directly penalise model developers”
AI Security Institute (renaming) (Regulations.ai, 2026-09) tier 4 press (aggregator)confirmed UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed2026-09-15
Publication rights 2/4
Anchor 2: Publishes; lab reviews with broad redaction.
evidence: tier 4 (press)
mechanism: statutory
Capped at 2 by ukaisi.04 (R.4): Most per-model results stay confidential; the public sees aggregated trends.. Floors up to 2 from ukaisi.03 do not exceed the cap.
- Found universal jailbreaks in every system tested and published that in the Frontier AI Trends Report. floors at 2 by R.4“AISI found universal jailbreaks across every tested system”
UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed2026-09-14 - Most per-model results stay confidential; the public sees aggregated trends. caps at 2 by R.4binding“evaluation work that would otherwise remain confidential between the Institute and individual AI labs”
UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed2026-09-14
Method transparency 3/4
Anchor 3: Tasks or code partly open.
evidence: tier 3 (self)
Floored at 3 by ukaisi.02 (M.1): Open-source Inspect framework and published methodology.. No admissible signal caps it.
- Open-source Inspect framework and published methodology. floors at 3 by M.1binding“An open-source framework for large language model evaluations”
UK AISI's Frontier AI Trends Report: Security Implications (Cloud Security Alliance, 2026-07-12) tier 4 press (primary)confirmed Inspect: an open-source framework for large language model evaluations (UK AI Security Institute, 2026) tier 3 self (ukaisi)confirmed2026-09-14
Role incompatibility 4/4
Anchor 4: No commercial products.
evidence: tier 3 (self)
Floored at 4 by ukaisi.10 (X.7): No commercial products.. No admissible signal caps it.
- No commercial products. floors at 4 by X.7binding“The Institute is an advisory technical research body rather than a formal regulator”
AI Security Institute (renaming) (Regulations.ai, 2026-09) tier 4 press (aggregator)confirmed About the AI Security Institute (UK AI Security Institute, 2026) tier 3 self (ukaisi)confirmed2026-09-14
Values under each evidence policy
| Dimension | Leads included | Retrieved & confirmed | Against interest | Verified spans | Primary only |
|---|---|---|---|---|---|
| Funding | 3 | 3 | 3 | 3 | – |
| Governance | 3 | 3 | 3 | 3 | 3 |
| Personnel | 3 | 3 | 3 | 3 | 3 |
| Access depth (lab-granted) | 3 | 3 | 3 | 3 | – |
| Scope control | 3 | 3 | 3 | 3 | – |
| Publication rights | 2 | 2 | 2 | 2 | – |
| Method transparency | 3 | 3 | 3 | 3 | – |
| Role incompatibility | 4 | 4 | 4 | 4 | – |
| Score, Lab procurement | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 2/8 |
| Score, Regulator or auditor | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 2/8 |
| Score, Public trust | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 2/8 |
| Score, Equal weights | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 8/8 | – No floor triggered, 2/8 |
Traced money and ties
hop 0: 2 rows (grant $7.5M); public: 1 row. Confirmed rows: 2 of 3 Dollar sums: confirmed + unaudited USD rows only; imported figures are not re-derived and never summed. Second hop traced for 2 of 2 sources. Ties: none within two steps recorded. Checked and not found: 1.
Ledger
Money in
- T45 from UK government (DSIT): grant undisclosed, FY. core budget about GBP 66m per financial year confirmedtier 3 selfhttps://www.aisi.gov.uk/about
- T46 from OpenAI: grant $7.50M, 2025. GBP 5.6m into the Alignment Project grant pool administered by AISI, alongside Anthropic, Microsoft, AWS and others confirmedtier 3 selfhttps://openai.com/index/advancing-independent-research-ai-alignment/
- T78 from OpenAI: commitment undisclosed, 2026-02. GBP 5.6m into the Alignment Project (£27m total); GBP not converted unauditedtier 3 selfhttps://www.aisi.gov.uk/blog/funding-60-projects-to-advance-ai-alignment-research
Checked and not found
- N08 Does Coefficient Giving fund the UK AI Security Institute?
Searched Coefficient grants index, 2,911 rows, snapshot 2026-09-11: no grant from Coefficient Giving in its index. Prompted by imported lead: kevinnbass ledger row EV11. importedhttps://coefficientgiving.org/grants/
Roles hosted
- R61 OpenAI: pays, 2025. into the Alignment Project pool; imported kevinnbass/metr-money-figure:EV11 importedhttps://www.aisi.gov.uk/grants
- R62 Anthropic: pays, 2025. as above importedhttps://www.aisi.gov.uk/grants
- R85 Halcyon Futures: pays, 2026. research pass 2026-09-15; Alignment Project backer unauditedhttps://alignmentproject.aisi.gov.uk/about
- R86 Sympatico Ventures: pays, 2026. research pass 2026-09-15; Alignment Project backer unauditedhttps://alignmentproject.aisi.gov.uk/about
- R87 Microsoft: pays, 2026. research pass 2026-09-15; Alignment Project backer unauditedhttps://alignmentproject.aisi.gov.uk/about
- R88 Amazon: pays, 2026. research pass 2026-09-15; AWS, Alignment Project backer unauditedhttps://alignmentproject.aisi.gov.uk/about
- R89 Schmidt Sciences: pays, 2026. research pass 2026-09-15; Alignment Project backer unauditedhttps://alignmentproject.aisi.gov.uk/about
- R90 Renaissance Philanthropy: contractor, 2025. research pass 2026-09-15; administers the Alignment Project unauditedhttps://alignmentproject.aisi.gov.uk/about
Funding graph, focused
Every one-way chain to a lab and to an evaluator at full strength, everything else faded. Hover any name to trace its connections; click a name to open its page.