DataPacts

Gray Swan AI: Red-Teaming Arena, Adversarial Evaluations and Attack Data for AI Labs

Checked

At a glance

Gray Swan AI, the brand of Gray Swan Security Inc., is a US AI security company that runs adversarial evaluations for frontier AI labs and a crowdsourced red-teaming Arena that, by its own count, has produced more than 1,000,000 attack trajectories as of May 28, 2026.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Arena participants produce attack trajectories with supply terms not published; Gray Swan’s own adversarial evaluations go to labs; labs named as users are Gray Swan’s claim; no license or data program on record.

It states a USD 40 million Series A of the same date and more than 20 customers, three of them named labs, while no license, sale or data program involving Gray Swan is on record as of 2026-11-09.

Every figure on this page is Gray Swan's own statement or an SEC filing, each with its date, read in the context of AI training data licensing. DataPacts lists no transaction that a party has not announced.

Each value names its source and check date, and how DataPacts sources and dates every figure is set out in the methodology.

What Gray Swan AI Sells to AI Labs

Gray Swan sells AI security work in three forms that grew out of Carnegie Mellon University research: Shade, a red-teaming agent; Cygnal, runtime protection for deployed models; and the Arena, where outside participants attack models. That is Gray Swan's own description, checked 2026-11-09.

For frontier labs, Gray Swan states that it runs adversarial evaluations, meaning tests that probe a model with inputs designed to make it fail.

The Arena is a second source of material: the attack data the Arena produces comes from outside participants, while the evaluations Gray Swan runs for labs are its own service. The sources checked do not say how the two connect, and no price for either is on record.

No terms for supplying attack data to a customer are published in the sources checked, and no Gray Swan program for licensing anyone's data appears among the programs tracked. The hub places Gray Swan under the safety and red-teaming kind among the types of RL environments, where it is a security business first.

Gray Swan AI Company Identity and SEC Filings

Gray Swan AI's legal name is Gray Swan Security Inc., a Delaware corporation registered with the SEC under CIK 0002015570 with a Pittsburgh, Pennsylvania address. The table lists each identity attribute with its source, as of 2026-11-09.

Attribute Value Source
Legal name Gray Swan Security Inc. grayswan.ai/news/gray-swan-announces-series-a
Brand Gray Swan AI grayswan.ai
Website https://www.grayswan.ai/ grayswan.ai
SEC registrant CIK 0002015570 sec.gov (EDGAR)
Form D filings 2024-03-21 and 2024-05-24 sec.gov (EDGAR)
State of incorporation Delaware sec.gov (EDGAR)
City Pittsburgh, PA (per Form D) sec.gov (EDGAR)
Country United States sec.gov (EDGAR)
Founded not stated on a primary source checked n/a
Type AI security company, listed as an RL environment builder grayswan.ai

The founding year, the names of the founders and a street address are not stated on a primary source checked, so this page gives none.

Secondary profiles disagree on the year, and DataPacts does not take a year, names or titles from them. The identity rows above therefore come only from Gray Swan's announcement and the SEC record, and each is dated by the check of 2026-11-09.

The Gray Swan Arena and Its Attack Trajectories

The Gray Swan Arena is a crowdsourced red-teaming venue where outside participants try to break AI models, and Gray Swan states that more than 15,000 participants have produced more than 1,000,000 attack trajectories there, as of May 28, 2026. The table sets the two figures beside what the record does not show, as of 2026-11-09.

Item Value Stated by Date Source
Participants more than 15,000 Gray Swan 2026-05-28 grayswan.ai/news/gray-swan-announces-series-a
Attack trajectories more than 1,000,000 Gray Swan 2026-05-28 grayswan.ai/news/gray-swan-announces-series-a
Rules and prize terms not on record in the sources checked n/a 2026-11-09 n/a
Ownership and supply of submissions not published in the sources checked n/a 2026-11-09 n/a

Gray Swan's own figures; no third party has counted them.

An attack trajectory is the recorded sequence of prompts and model responses in one attempt to break a model. Trajectories in general, and who buys agent trajectory data, are covered on their own page, which lists Gray Swan's count as a trajectory source that produces its own data rather than licensing it from holders.

The record does not show who owns a participant's submission, whether customers receive raw trajectories or only evaluation results, or on what license either would pass. This page is updated when Gray Swan publishes terms, and a reader who needs them before then has to ask Gray Swan directly.

Taking Part in the Arena: What Participants Can Check

Gray Swan's Arena rules and prize terms are not on record in the sources DataPacts checks, so this page states no prize amount, payment method or eligibility rule.

A would-be participant can answer five questions from the Arena's own current rules before submitting anything: who owns a submitted attack, whether it can reach a customer's training or evaluation work, how and when prizes are paid, which countries are eligible, and what the Arena does with the participant's account data. This page gives no answer to any of them.

Paid task work for AI labs in general, with rates and assessments, is compared under AI training jobs and platforms.

Labs Gray Swan AI Names as Users

Gray Swan states that it has more than 20 customers across frontier labs and enterprises and names Anthropic, OpenAI and Meta as users, in its Series A announcement of May 28, 2026.

It adds that 11 model system cards cite Gray Swan, a count that is Gray Swan's own, and no system card is quoted on this page. The table lists each claim, as of 2026-11-09.

Claim Value Stated by Date Source
Customers over 20, across frontier labs and enterprises Gray Swan 2026-05-28 grayswan.ai/news/gray-swan-announces-series-a
Labs named as users Anthropic, OpenAI, Meta Gray Swan 2026-05-28 grayswan.ai/news/gray-swan-announces-series-a
Model system cards citing Gray Swan 11 Gray Swan 2026-05-28 grayswan.ai/news/gray-swan-announces-series-a
License, deal or data program on record none DataPacts check 2026-11-09 n/a

A named user is not a published license or data transfer.

"Named as users" is a customer claim, not a published license, a purchase price or a data transfer, and no deal or data program naming Gray Swan is on record as of 2026-11-09. The sources checked contain no statement from Anthropic, OpenAI or Meta on the claim, so the page repeats it as Gray Swan's and adds nothing to it.

The same reasoning applies to the count of customers: it is a number Gray Swan gives without a list.

Labs' own data programs and terms are compared in the directory of who buys AI training data.

Gray Swan AI Series A, as Gray Swan States It

Gray Swan states that it raised a USD 40 million Series A on May 28, 2026, co-led by Wing Venture Capital and Madrona. The table sets the round beside the two earlier SEC filings, as of 2026-11-09.

Date Event Amount Lead investors Stated by Source
2024-03-21 Form D filed by Gray Swan Security Inc. not recorded not recorded SEC EDGAR sec.gov
2024-05-24 Form D filed not recorded not recorded SEC EDGAR sec.gov
2026-05-28 Series A USD 40,000,000 Wing Venture Capital, Madrona (co-led) Gray Swan grayswan.ai

Two Form D filings by Gray Swan Security Inc. are dated 2024-03-21 and 2024-05-24; their amounts are not recorded here, and no valuation is stated by Gray Swan in the sources checked.

No other total is stated on this page, because the only amount Gray Swan itself states is the USD 40 million round.

Funding is what investors paid for shares, not a price for attack data and not a measure of revenue, and rounds across AI data companies are compared under funding and valuations of AI data companies.

Working With Gray Swan AI: Benefits, Risks, Downsides and Alternatives

For a red-teamer, a data holder or an environment builder, Gray Swan is a security vendor with a public Arena and an SEC-registered entity but no published data terms: its figures are its own claims, and its prize and license terms are not on record.

Benefits

  • An SEC-registered entity with two Form D filings, dated 2024-03-21 and 2024-05-24.
  • A public Arena at a stated scale of more than 15,000 participants and more than 1,000,000 attack trajectories (Gray Swan, 2026-05-28).
  • Three labs named as users: Anthropic, OpenAI and Meta (Gray Swan's claim).

Risks

  • No published terms on who owns a submitted attack or on how attack data reaches customers.
  • Prize terms are not on record, so a participant cannot check what is offered from this page.
  • Every scale and customer figure is Gray Swan's own, and none has an outside count.

Downsides

  • No program for licensing company data or buying data from holders is on record as of 2026-11-09.
  • Gray Swan is not a general RL-environment vendor, so builders of other environment types gain little from it.

Alternatives

  • Anthropic's model safety bug bounty, announced 2024-08-08, where Anthropic stated rewards of up to USD 15,000 for novel universal jailbreak attacks, by invitation through HackerOne; its current status is not checked.
  • Contributor task platforms such as Shipd, which shows bounties of USD 80 to USD 220 per item (checked 2026-11-09) and whose terms assign all work product to Shipd, operated by Datacurve AI, Inc. (Contributor Terms s. 3.1).
  • Building a product directly: routes to sell an RL environment or task set are compared separately.
  • Not taking part or licensing at all, which is a valid outcome.

Gray Swan's figures are its own and can change, so the Arena's current rules and Gray Swan's terms are the documents to read before relying on any figure here.

Gray Swan AI Among Safety and Evaluation Builders

Gray Swan is one of two evaluation and safety companies on DataPacts' list of RL environment builders, kept there because its own records show attack data rather than benchmarks alone. The full list is under RL environment companies.

More About Gray Swan AI

Frequently Asked Questions About Gray Swan AI

Who founded Gray Swan AI?

Gray Swan describes itself as coming out of Carnegie Mellon University research, and the sources checked name no founder on a primary page. No founder is named here.

Where is Gray Swan AI based?

Gray Swan's SEC Form D filings give Pittsburgh, Pennsylvania. The entity, Gray Swan Security Inc., is a Delaware corporation.

How much funding has Gray Swan AI raised?

Gray Swan states a USD 40 million Series A on 2026-05-28, co-led by Wing Venture Capital and Madrona. No total across rounds is stated here.

Does Gray Swan pay Arena participants?

The Arena's rules and prize terms are not on record in the sources checked. Read the Arena's current rules before taking part.

Is Gray Swan AI an RL environment company?

It is an AI security company, not a general RL-environment vendor. It is listed among builders because its Arena produces attack trajectories.

Does Gray Swan AI buy company data?

No program for licensing or buying company data is on record as of 2026-11-09.