Methodology

This site lists 23 AI red teaming vendors alphabetically, with facts taken only from pages that were opened and are linked on each vendor page. Nothing was tested, rated or scored. Last full check: 2026-11-07.

What counts as a vendor here

A company is listed if, on the check date, all of these held:

• It had a live, public page describing an AI red teaming product or service it sells to outside buyers (not only an internal research team).
• That page, or a linked data sheet, pricing page or docs page, could be opened and read.
• The offering tests AI systems themselves (models, LLM apps, agents), not only the surrounding infrastructure.

Companies whose standalone brand no longer exists are listed under the acquirer's current product name: Robust Intelligence as Cisco AI Defense, Protect AI as Palo Alto Networks Prisma AIRS, CalypsoAI as F5 AI Red Team.

Who was left out, and why

• Scale AI: no live red-teaming product page (scale.com/red-teaming returns 404). Red teaming appears only as a capability claim in its GenAI platform FAQ, and the offering is aimed mainly at model developers and government rather than enterprises deploying LLM apps.
• Haize Labs: acquired by Beacon Software (Toronto) and turned into Beacon's in-house "Applied AI Research Group". haizelabs.com returned HTTP 403 on every request and sales to outside buyers could not be confirmed. Sources also disagree on the deal date.
• Not researched: Prompt Security (SentinelOne), Aim Security (Cato Networks), WitnessAI, Troj.ai and Virtue AI. No pages were opened for them, so nothing is claimed about them. They may be added in a later check.

How facts were gathered

• Vendor facts (offering type, what it tests, framework mappings, pricing, free tier, headquarters) come from the vendor's own product, pricing, docs or about pages. They are recorded as claims, not verified capabilities.
• Ownership comes from the acquirer's press release or results statement where one exists, with the date the source gives. Where only an announcement was found and no closing notice, the entry says "announced". No acquisition price is published on this site because the figures found were press reports that were not independently confirmed.
• Standards and regulation come from OWASP, MITRE, NIST, the EU AI Act text and timeline, the Official Journal summary by a law publisher, and the White House.
• Buyer guidance (cadence, humans vs automation, deliverables) comes from Microsoft's published red-teaming guidance and its AI Red Team's paper, plus vendor pages.
• Most vendor pages were read with automated fetch tools and then reviewed. Vendor sites change often; the linked source is always the authority over this site.

What this site deliberately does not do

• No ranking, rating, score or "top 10" order. Vendors are alphabetical within offering type.
• No hands-on testing of any vendor. Nothing here says "we tested".
• No customer quotes, case-study results, or market-share figures.
• No prices other than those printed on a vendor's own pricing page.
• No funding or valuation figures other than rounds the vendor itself announced.
• No ownership claim where current ownership could not be confirmed on an opened page (this applies to NCC Group).

Known uncertainties

These were flagged during research and are carried onto the relevant vendor pages:

• Promptfoo and OpenAI: the deal was announced 2026-03-09 and Promptfoo's own pages now say it is part of OpenAI, but no closing date was found.
• SPLX and Zscaler: Zscaler's release says "has acquired" while its boilerplate says "proposed acquisition"; no separate closing notice was found.
• Pillar Security: deal databases list a December 2025 acquisition of a different company with the same name (an Ontario security-guard firm). The AI vendor is independent.
• HiddenLayer mappings: search results suggested its AI Discovery module maps to MITRE ATLAS, OWASP and NIST; the red-teaming pages opened claim none, so none is shown.
• Enkrypt AI pricing: the pricing page contradicts itself on which paid plans include multimodal red teaming.
• Gray Swan Arena size: the site says 23,000+ researchers; a May 2026 press article says 15,000+. Neither figure is used on this site.
• "MITRE" vs "MITRE ATLAS": Lasso and Repello say "MITRE" without naming ATLAS; they are recorded as "MITRE (ATLAS not named)".
• OWASP edition: HackerOne cites an "OWASP LLM Top 10 (2026)". OWASP's project page lists 2025 as the current LLM edition; this site uses "OWASP Top 10 for LLM Applications (2025)".
• Headquarters: not confirmed on an opened page for Lasso, Holistic AI, Repello, Adversa, SPLX, Prisma AIRS, HackerOne, NCC Group, Noma and Straiker. Trail of Bits shows only a New York mailing address.
• NIST CAISI naming: the NIST page rendered the center's name inconsistently; this site refers to it as "NIST's AI standards center (CAISI)".

Update policy

• Every vendor page and the standards page are re-checked against their sources at least once a quarter, and the "Last updated" date is changed only when a check happens.
• An acquisition, a pricing change or a product withdrawal triggers an out-of-cycle update to that vendor's page.
• Corrections from vendors or readers are checked against the vendor's own pages and applied if they hold; a vendor cannot pay to change, add or remove an entry.
• Vendors not yet researched are added only after their pages have been opened and read under the same rules.

Corrections

Send corrections with a link to the page that proves them to CONTACT_EMAIL. See also the disclosure for how this site makes money and what that does and does not affect.