Center for AI Safety
Scorecard
Hazard benchmarks (WMDP, HLE); director advises xAI.
Nonprofit, San Francisco, US. Independent referee; listed with independent referees. Confidence med. Domains: benchmarks, misuse.
Conditional floor 8/8 evidenced · Retrieved & confirmed Scores by preset: Lab procurement –; Regulator or auditor –; Public trust –; Equal weights –. Weakest evidenced dimension: access depth (lab-granted) 0/4. At least one evidenced conflict dimension (funding, governance, personnel, role incompatibility, scope, publication) at 1: usable with conditions the card names.
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. A public recusal policy covering the xAI relationship.
- Funding 3 to 4: score –
- Governance 2 to 3: score –
- Personnel 1 to 2: score –, band becomes no floor triggered
Dissent, lower. Personnel should be 0. The director advises xAI while CAIS benchmarks (HLE, WMDP, MASK) are used to grade xAI models, and no recusal statement, disclosure page, or conflict policy is published anywhere on safe.ai. P.1 sets 0 for a governance role without recusal; an unrecused standing advisory role at a graded lab, with the benchmark scoring under the same director, is the nearest thing to it.
Dissent, higher. Personnel should be 2. An advisory role is not a board or committee seat, so P.2's cap of 1 rather than P.1's 0 applies, and nothing above 1 was ever claimed. HLE and WMDP are scored mechanically on public leaderboards where xAI's results are visible to anyone, no equity, hiring flow, or second lab tie is documented, and a published recusal note would lift the value to 2.
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.
- Funded mainly by SFF rather than lab-linked pools. floors at 3 by F.6“SFF gave $1.1m to CAIS and $1.6m to the action fund”
It looks like there are some good funding opportunities in AI safety right now (80,000 Hours, 2025-10-13) tier 4 press (primary)confirmed2026-09-14 - SFF recommended $289,000 to CAIS in 2025 (about $1.1M cumulative per 80,000 Hours); SFF's funder Jaan Tallinn led Anthropic's Series A. caps at 3 by F.6binding“Center for AI Safety (CAIS) Main: $210,000 Freedom: $78,000 Fairness: $0 Mean: $0 $289,000”
SFF-2025 S-Process Recommendations Announcement (Survival and Flourishing Fund, 2025) tier 3 self (sff)confirmed Anthropic raises $124 million Series A (Anthropic, 2021-05-28) tier 3 self (anthropic)confirmed It looks like there are some good funding opportunities in AI safety right now (80,000 Hours, 2025-10-13) tier 4 press (primary)confirmed2026-09-15
- Nonprofit. floors at 2 by G.1binding“is a San Francisco-based research and field-building nonprofit”
Center for AI Safety (CAIS, 2026) tier 3 self (cais)confirmed2026-09-14
- Director is a safety adviser to xAI while the organization's benchmarks are used to grade xAI models. caps at 1 by P.2binding“have driven some of the bigger successes in AI policy and advise xAI”
It looks like there are some good funding opportunities in AI safety right now (80,000 Hours, 2025-10-13) tier 4 press (primary)confirmed Center for AI Safety (CAIS, 2026) tier 3 self (cais)confirmed2026-09-14 - Document request: Request to the Center for AI Safety: the terms of the director's advisory role at xAI (start date, compensation or equity if any) and any recusal or disclosure statement covering CAIS benchmark results on xAI models.
- Public-model access. caps at 0 by A.6binding“We evaluate a range of leading LLMs”
Center for AI Safety (CAIS, 2026) tier 3 self (cais)confirmed Research Projects | CAIS (Center for AI Safety) tier 3 self (cais)confirmed2026-09-14
- Sets own benchmark design. floors at 3 by S.5binding“CAIS develops foundational benchmarks and methods which concretize the problem”
Center for AI Safety (CAIS, 2026) tier 3 self (cais)confirmed2026-09-14
- Publishes results. floors at 2 by R.1binding“The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems Mar 5, 2025”
Center for AI Safety (CAIS, 2026) tier 3 self (cais)confirmed2026-09-14
- Open benchmarks widely used. floors at 3 by M.3binding“CAIS develops foundational benchmarks and methods which concretize the problem”
Center for AI Safety (CAIS, 2026) tier 3 self (cais)confirmed2026-09-14
- Co-produced Humanity's Last Exam with Scale AI, a Meta-owned vendor. caps at 3 by X.6binding“1 Center for AI Safety, 2 Scale AI”
Center for AI Safety (CAIS, 2026) tier 3 self (cais)confirmed Humanity's Last Exam (Center for AI Safety and Scale AI, 2025) tier 3 self (cais)confirmed Scale AI not winding down following Meta deal, interim CEO says (CNBC, 2025-06-18) tier 4 press (primary)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 | 2 | 2 | – | 2 | – |
| Personnel | 1 | 1 | 1 | 1 | – |
| Access depth (lab-granted) | 0 | 0 | 0 | 0 | – |
| Scope control | 3 | 3 | – | 3 | – |
| Publication rights | 2 | 2 | – | 2 | – |
| Method transparency | 3 | 3 | – | 3 | – |
| Role incompatibility | 3 | 3 | 3 | 3 | – |
| Score, Lab procurement | – Conditional floor, 8/8 | – Conditional floor, 8/8 | – Conditional floor, 4/8 | – Conditional floor, 8/8 | – Unevidenced, 0/8 |
| Score, Regulator or auditor | – Conditional floor, 8/8 | – Conditional floor, 8/8 | – Conditional floor, 4/8 | – Conditional floor, 8/8 | – Unevidenced, 0/8 |
| Score, Public trust | – Conditional floor, 8/8 | – Conditional floor, 8/8 | – Conditional floor, 4/8 | – Conditional floor, 8/8 | – Unevidenced, 0/8 |
| Score, Equal weights | – Conditional floor, 8/8 | – Conditional floor, 8/8 | – Conditional floor, 4/8 | – Conditional floor, 8/8 | – Unevidenced, 0/8 |
Traced money and ties
hop 2: 2 rows (recommendation $1.4M). Confirmed rows: 2 of 2 Dollar sums: confirmed + unaudited USD rows only; imported figures are not re-derived and never summed. Second hop traced for 1 of 1 sources. Ties: none within two steps recorded. Checked and not found: 0.
Ledger
Money in
- T57 from Survival and Flourishing Fund: recommendation $1.10M, 2023-2025. SFF to CAIS, about $1.1M confirmedtier 4 presshttps://80000hours.org/2025/01/it-looks-like-there-are-some-good-funding-opportunities-in-ai-safety-right-now/
- T71 from Survival and Flourishing Fund: recommendation $0.29M, 2025. SFF 2025: $289,000 to CAIS; the CAIS Action Fund is a separate entity ($772,000, not added) confirmedtier 3 selfhttps://survivalandflourishing.fund/2025/recommendations
Roles hosted
- R43 Dan Hendrycks: principal. director confirmedhttps://safe.ai/
One hop away
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.