evaluator; distance from a lab: 1
Gray Swan
Scorecard
Crowdsourced adversarial testing (Arena) plus Shade and Cygnal defense products; cited in 11 recent system cards.
VC-backed, Pittsburgh, US. Vendor / red-team co.; listed with commercial and first-party. Confidence high. Domains: jailbreak, cyber, misuse.
8/8 evidenced · Retrieved & confirmed Scores by preset: Lab procurement –; Regulator or auditor –; Public trust –; Equal weights –. Weakest evidenced dimension: role incompatibility 0/4. At least one evidenced conflict dimension at 0: not a candidate for an independence-critical role until the floor moves. Access and methods never set a band.
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. Structural separation of the evaluation business, and independent governance for the OpenAI-related conflict beyond personal recusal.
- Funding 1 to 2: score –
- Governance 1 to 2: score –
- Personnel 1 to 2: score –
Dissent, lower. Personnel (1 to 0). The co-founder and chief scientist chairs OpenAI's Safety and Security Committee, which oversees the very releases in which Gray Swan is cited. The only evidence of recusal is one 2024 press sentence describing a personal arrangement; no written policy, scope, or administrator is public. P.1 gives 0 where a recusal is not documented, and a single unverifiable press line is thin documentation.
Dissent, higher. Role incompatibility (0 to 1). The record documents that labs use Gray Swan's platform and are named in its system-card citations, but not that any lab buys Cygnal, the runtime defense product; lab use may be limited to Arena and Shade, which are testing services. X.1's 0 requires a defense or monitoring product sold to an evaluated lab; on the documented facts, anchor 1 (services to labs) fits.
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 1/4
Anchor 1: Material revenue or investment from evaluated labs or their investors.
evidence: tier 3 (self)
Capped at 1 by grayswan.03 (F.5): $40M Series A in May 2026 at a $200M valuation; revenue from every major lab.. Floors up to 1 from grayswan.06 do not exceed the cap.
- $40M Series A in May 2026 at a $200M valuation; revenue from every major lab. caps at 1 by F.5binding“Gray Swan, The AI Security Company Trusted by Every Major Frontier Lab, Raises $40M Series A”
Gray Swan announces Series A (Gray Swan, 2026-05-28) tier 3 self (grayswan)confirmed This AI Startup's Army of 15,000 Hackers (Forbes, 2026-05-28) tier 4 press (primary)confirmed2026-09-14 - Breadth across OpenAI, Anthropic, Google DeepMind, Meta, xAI and ByteDance, so no single lab dominates. floors at 1 by F.5 (anchor text)
This AI Startup's Army of 15,000 Hackers (Forbes, 2026-05-28) tier 4 press (primary)confirmed2026-09-14
Governance 1/4
Anchor 1: Private for-profit with no published COI policy.
evidence: tier 3 (self)
Capped at 1 by grayswan.05 (G.2): Venture-backed for-profit; no published COI policy found..
- Venture-backed for-profit; no published COI policy found. caps at 1 by G.2binding“today announced a $40 million Series A round co-led by Wing Venture Capital and Madrona”
Gray Swan announces Series A (Gray Swan, 2026-05-28) tier 3 self (grayswan)confirmed2026-09-14 - Series A co-led by Wing and Madrona with Obvious Ventures, Snowflake Ventures, Hudson River Trading, Samsung Next and Magarac; the co-lead's own post notes the chief scientist serves on OpenAI's board. caps at 2 by G.2 (anchor text); G.3 not triggered“serves on the board of OpenAI”
Why We Are Co-Leading Gray Swan's Series A (Madrona, 2026-05-28) tier 4 press (primary)confirmed Gray Swan raises $40M Series A to secure frontier AI (Dealroom, 2026-06) tier 4 press (aggregator)unaudited2026-05-28 - Snowflake, a Series A co-investor, announced a commercial agreement with Anthropic rather than an equity stake, so the co-investor is not a lab investor under F.4 or G.3. informational (Informational: closes the F.4/G.3 question for one co-investor; no floor follows from the absence of a tie (RULES 0).)
Snowflake and Anthropic Announce $200 Million Partnership to Bring Agentic AI to Global Enterprises (Snowflake, 2025-12-03) tier 3 self (snowflake-ventures)confirmed2026-09-15
Personnel 1/4
Anchor 1: Leaders hold equity or advisory roles at labs; informal recusal.
evidence: tier 3 (self)
Capped at 1 by grayswan.01 (P.1): Co-founder and chief scientist chairs OpenAI's Safety and Security Committee and sits on its nonprofit board.; capped at 1 by grayswan.12 (P.1): Chief scientist appointed to OpenAI's board in 2024 and chairs its safety and security committee; also a 2025 Schmidt Sciences AI safety funding recipient.. Floors up to 1 from grayswan.02 do not exceed the cap.
- Co-founder and chief scientist chairs OpenAI's Safety and Security Committee and sits on its nonprofit board. caps at 1 by P.1binding“he chairs the Safety and Security Committee”
Zico Kolter Joins OpenAI's Board of Directors (OpenAI, 2024-08) tier 3 self (openai)confirmed Bio (Zico Kolter, 2026) tier 3 self (grayswan)confirmed2026-09-14 - The chief scientist recuses himself from Gray Swan's dealings with OpenAI. floors at 1 by P.1“he has recused himself from interactions between the two companies”
Gray Swan AI Is Working With OpenAI to Red Team Its Models (Forbes, 2024-10-29) tier 4 press (primary)confirmed This hacker team is bulletproofing AI models for companies like OpenAI and Anthropic (Forbes Australia, 2024-11-07) tier 4 press (primary)confirmed2026-09-14 - Chief scientist appointed to OpenAI's board in 2024 and chairs its safety and security committee; also a 2025 Schmidt Sciences AI safety funding recipient. caps at 1 by P.1binding“became chair of its safety and security committee”
Zico Kolter (Wikipedia, 2026) tier 4 press (aggregator)confirmed2026-09-15 - Document request: Request from Gray Swan a copy or description of the recusal arrangement governing the chief scientist's role between Gray Swan and OpenAI's board and Safety and Security Committee (date adopted, scope, who administers it), as reported by Forbes on 29 October 2024. Per RULES 12 this goes to the person before publication.
- Document request: Request from Gray Swan (or from the funds) a statement of whether Wing, Madrona, Obvious Ventures, Snowflake Ventures, Hudson River Trading, Samsung Next or Magarac Venture Partners hold equity in a frontier developer. The public record shows Snowflake's Anthropic tie as a $200M commercial partnership (3 Dec 2025), not equity.
Access depth (lab-granted) 2/4
Anchor 2: Pre-release with safeguards off or extended time.
evidence: tier 3 (self)
mechanism: lab-controlled
Floored at 2 by grayswan.07 (A.1): Embedded in pre-release safety evaluation processes; cited in 11 system cards.. No admissible signal caps it.
- Embedded in pre-release safety evaluation processes; cited in 11 system cards. floors at 2 by A.1binding“Gray Swan’s benchmarks are embedded into the safety evaluation processes”
Gray Swan announces Series A (Gray Swan, 2026-05-28) tier 3 self (grayswan)confirmed Why We Are Co-Leading Gray Swan's Series A (Madrona, 2026-05-28) tier 4 press (primary)confirmed2026-09-14
Scope control 2/4
Anchor 2: Scope negotiated per engagement.
evidence: tier 3 (self)
Capped at 2 by grayswan.10 (S.2): Scope set by lab engagements..
- Scope set by lab engagements. caps at 2 by S.2binding“AI Red-Teaming Dedicated adversarial assessment, scoped to your deployment”
Gray Swan announces Series A (Gray Swan, 2026-05-28) tier 3 self (grayswan)confirmed2026-09-14
Publication rights 2/4
Anchor 2: Publishes; lab reviews with broad redaction.
evidence: tier 3 (self)
mechanism: lab-controlled
Capped at 2 by grayswan.09 (R.3): Findings appear as lab-summarized system-card citations rather than independent reports..
- Findings appear as lab-summarized system-card citations rather than independent reports. caps at 2 by R.3binding“Gray Swan is a leading security evaluator cited in 11 recent frontier model system cards”
Gray Swan announces Series A (Gray Swan, 2026-05-28) tier 3 self (grayswan)confirmed2026-09-14
Method transparency 2/4
Anchor 2: Methods described in prose.
evidence: tier 3 (self)
Floored at 2 by grayswan.08 (M.2): Founders wrote foundational jailbreak papers; Arena results partly public.. No admissible signal caps it.
- Founders wrote foundational jailbreak papers; Arena results partly public. floors at 2 by M.2binding“Research Published papers and findings from our research team”
This AI Startup's Army of 15,000 Hackers (Forbes, 2026-05-28) tier 4 press (primary)confirmed Gray Swan announces Series A (Gray Swan, 2026-05-28) tier 3 self (grayswan)confirmed2026-09-14
Role incompatibility 0/4
Anchor 0: Sells defense or monitoring products to evaluated labs.
evidence: tier 3 (self)
held at a supportable anchor
Capped at 0 by grayswan.04 (X.1): Sells Shade and Cygnal, the fixes, to the same labs whose models it evaluates..
- Sells Shade and Cygnal, the fixes, to the same labs whose models it evaluates. caps at 0 by X.1binding“Cygnal for real-time AI protection, Shade for continuous adversarial testing”
Gray Swan announces Series A (Gray Swan, 2026-05-28) tier 3 self (grayswan)confirmed Why We Are Co-Leading Gray Swan's Series A (Madrona, 2026-05-28) tier 4 press (primary)confirmed2026-09-14 - A podcast interview reports Shade, Gray Swan's defense product, in use at Anthropic, a lab whose models Gray Swan red-teams. caps at 0 by X.1
Red-Teaming after Mythos — Zico Kolter & Matt Fredrikson, Gray Swan (Latent Space, 2026-06-22) tier 4 press (aggregator)confirmed2026-09-15
Values under each evidence policy
| Dimension | Leads included | Retrieved & confirmed | Against interest | Verified spans | Primary only |
|---|---|---|---|---|---|
| Funding | 1 | 1 | 1 | 1 | – |
| Governance | 1 | 1 | 1 | 1 | – |
| Personnel | 1 | 1 | 1 | 1 | – |
| Access depth (lab-granted) | 2 | 2 | 2 | 2 | – |
| Scope control | 2 | 2 | 2 | 2 | – |
| Publication rights | 2 | 2 | 2 | 2 | – |
| Method transparency | 2 | 2 | 2 | 2 | – |
| Role incompatibility | 0 | 0 | 0 | 0 | – |
| Score, Lab procurement | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Unevidenced, 0/8 |
| Score, Regulator or auditor | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Unevidenced, 0/8 |
| Score, Public trust | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Unevidenced, 0/8 |
| Score, Equal weights | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Disqualifying floor, 8/8 | – Unevidenced, 0/8 |
Traced money and ties
hop 0: 1 row; unattributed: 1 row (investment $40.0M). Confirmed rows: 2 of 2 (1 excluded after a re-fetch disagreed or could not be verified: T81) Dollar sums: confirmed + unaudited USD rows only; imported figures are not re-derived and never summed. Second hop traced for 1 of 2 sources: untraced Wing Venture Capital. Ties: none within two steps recorded. Checked and not found: 0.
Ledger
Money in
- T21 from Wing Venture Capital: investment $40.00M; round total, not one investor's check, 2026-05. Series A co-led with Madrona; $200M valuation confirmedtier 3 selfhttps://www.grayswan.ai/news/gray-swan-announces-series-a
- T26 from Anthropic: contract undisclosed, 2024-2026. paid red-teaming and product use; cited in system cards confirmedtier 3 selfhttps://www.grayswan.ai/news/gray-swan-announces-series-a
- T81 from Wing Venture Capital: investment $40.00M; superseded by T21 — excluded from sums, 2026-05-28. Series A co-led by Wing and Madrona with Obvious, Snowflake Ventures, Hudson River Trading, Samsung Next, Magarac supersededtier 4 presshttps://www.madrona.com/gray-swan-series-a/
Roles hosted
- R27 Zico Kolter: principal. co-founder, chief scientist confirmedhttp://zkolter.github.io/bio/
- R77 Obvious Ventures: investor, 2026-05. research pass 2026-09-15 unauditedhttps://app.dealroom.co/news/note/gray-swan-raises-40m-series-a-to-secure-frontier-ai
- R78 Snowflake Ventures: investor, 2026-05. research pass 2026-09-15 unauditedhttps://app.dealroom.co/news/note/gray-swan-raises-40m-series-a-to-secure-frontier-ai
- R79 Hudson River Trading: investor, 2026-05. research pass 2026-09-15 unauditedhttps://app.dealroom.co/news/note/gray-swan-raises-40m-series-a-to-secure-frontier-ai
- R80 Samsung Next: investor, 2026-05. research pass 2026-09-15 unauditedhttps://app.dealroom.co/news/note/gray-swan-raises-40m-series-a-to-secure-frontier-ai
- R81 Magarac Venture Partners: investor, 2026-05. research pass 2026-09-15 unauditedhttps://app.dealroom.co/news/note/gray-swan-raises-40m-series-a-to-secure-frontier-ai
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.