Anthropic states a bounty of up to USD 15,000 per novel universal jailbreak (announced 2024-08-08, invite-only at launch), and Gray Swan states a public Arena with cash prizes and no amounts.
Red-teaming is the adversarial route within AI training data licensing: labs obtain attacks from contributors instead of licensing files that already exist. Every program term below is that program's own claim with its check date (2026-11-17 unless stated), and rates by platform sit on the pay page.
The pages read say nothing about exposure to harmful content for three platforms and nothing about rights in a submitted attack for any program.
Programs come from their own pages, and how we source, date and check every figure is set out in the methodology.
What red-teaming data is and why labs commission it
Red-teaming data is adversarial conversation data: a person attacks a model, the exchange is recorded and the result is scored.
The Anthropic authors' paper "Red Teaming Language Models to Reduce Harms" (arXiv 2209.07858, v1 2022-08-23, v2 2022-11-22) releases 38,961 red team attacks and reports that models trained with reinforcement learning from human feedback become harder to red team as they scale.
The stated purpose, in the paper's title and abstract, is to find and reduce harms.
Each attack serves two uses. As a test, a failed reply shows a weakness to fix before release. As training material, the same exchange, once rated, teaches the model to refuse, which is how the work connects to RLHF data, where ratings and rankings of replies are covered.
The paper releases its attacks as a dataset, and it is the only attack dataset read for this page.
Programs and platforms on record: who states what
Seven organizations are covered here by what their own pages state about red-teaming work, and for three of them the pages read state nothing that can be quoted.
Seven organizations and what their own pages state about red-teaming work, as of 2026-11-17 (DataAnnotation 2026-11-09), alphabetical, each term the organization's own claim.
| Organization | What its page states about red-teaming work | Pay as stated (claim) | Eligibility and access as stated | Checked | Source |
|---|---|---|---|---|---|
| Alignerr | no red-teaming role among the 60 listings visible on its jobs page | not stated | not stated for red-teaming | 2026-11-17 | alignerr.com/jobs |
| Anthropic | model safety bug bounty with HackerOne for novel universal jailbreaks; CBRN and cybersecurity focus; early access to a safety system not yet deployed; announced 2024-08-08, plans to expand | rewards up to USD 15,000 for novel universal jailbreak attacks | invite-only at launch; applications by 2024-08-16; experienced AI security researchers or people with proven jailbreak skills | 2026-11-17 | anthropic.com/news/model-safety-bug-bounty |
| DataAnnotation (Surge Labs) | job board of 133 roles in 14 categories with no category named red teaming or safety; the Cybersecurity category lists 4 roles and the page does not say whether they test models | not stated for red-teaming | not stated for red-teaming | 2026-11-09 | dataannotation.tech |
| Gray Swan (Gray Swan Security Inc.) | Arena, "the world's largest adversarial AI red teaming network"; homepage also states attacks by 23,000+ adversarial researchers and over three million attack attempts, not tied to the Arena; 15,000+ participants and 1,000,000+ attack trajectories (release of 2026-05-28) | "cash prizes", amounts not stated | no criteria stated on the homepage | 2026-11-17 | grayswan.ai |
| Handshake AI | no red-teaming role among the roles listed on its AI page; AI Evaluation Specialist listed | not stated for red-teaming; Evaluation Specialist up to USD 17/hr | US-based with valid work authorization; CPT/OPT may qualify, STEM OPT not supported | 2026-11-17 | joinhandshake.com/ai |
| OpenAI (Red Teaming Network) | page not read: HTTP 403 on 2026-11-17 | not stated | not stated | 2026-11-17 | openai.com/index/red-teaming-network |
| Scale AI | red-teaming page not read: HTTP 404 on scale.com/red-teaming on 2026-11-17 | not stated | not stated | 2026-11-17 | scale.com |
A stated row is the organization's claim on its check date, and "not stated" means the pages read say nothing, not that no program exists. Ranked lists of AI red-teaming software are a different market and out of scope.
Three ways red-teaming work is organized: bounties, arenas and paid tasks
Red-teaming work reaches contributors in three forms on the pages read: an invite-only bounty, a public arena with prizes, and paid conversation tasks.
The bounty is Anthropic's model safety bug bounty, run with HackerOne. Its announcement of 2024-08-08 offers "bounty rewards up to $15,000" for "novel, universal jailbreak attacks", with a focus on CBRN (chemical, biological, radiological and nuclear) and cybersecurity topics.
It was invite-only at launch with applications due by 2024-08-16, participants get early access to a safety mitigation system not yet deployed publicly, and the page says Anthropic plans to expand the initiative. The page shows no later status on 2026-11-17.
The arena is Gray Swan's. Its homepage on 2026-11-17 calls it "The world's largest adversarial AI red teaming network" with cash prizes, and elsewhere states attacks "discovered by 23,000+ adversarial researchers" and "over three million attack attempts" without tying either count to the Arena.
Its release of 2026-05-28 states more than 15,000 participants and more than 1,000,000 attack trajectories. Both are Gray Swan's claims in different wording and are shown side by side, not merged. Its legal entity and funding are on the profile of Gray Swan AI: red-teaming Arena and attack data.
Paid tasks are the oldest form in the record. The 2022 paper recruited 324 US-based crowdworkers, 307 through Amazon Mechanical Turk and 17 through Upwork, to write attacks as conversations.
How pay for red-teaming is stated, with dates
Pay for red-teaming is stated in three framings on the pages read, a per-finding bounty ceiling, unstated cash prizes and an hourly or per-task rate, and each is that program's own claim on its check date, not an earning.
Anthropic's "up to USD 15,000" (announcement 2024-08-08) is a ceiling per novel universal jailbreak, not an average and not a current offer.
Gray Swan's prizes carry no amount on the homepage read on 2026-11-17. The 2022 paper is a historic baseline: USD 7.50 to 9.50 per set of five conversations on Mechanical Turk and USD 20 per hour on Upwork.
No red-teaming rate is on record for Alignerr, DataAnnotation, Handshake AI or Outlier. Handshake AI's listed Evaluation Specialist, up to USD 17 per hour on 2026-11-17, is evaluation work and is not offered as a red-teaming rate. Every dated rate is on the page for pay for AI training work by platform and field.
Exposure to harmful content: what primary sources say
Red-teaming asks people to write prompts about harmful subjects, and the two primary sources read for this page, an AI developer's own paper and a court docket, treat the exposure differently: the paper reports precautions and a survey, the docket records an allegation.
In the Anthropic authors' paper, red team members were given warnings that they might be exposed to sensitive content, had no required topics, and reviewers could preview a transcript and skip it.
The authors report that surveys and informal feedback found red team members "enjoyed the task and did not experience significant negative emotions". That is the authors' statement about their own 2022 crowdworker study, not a finding about any other program.
The Handshake AI page, the Alignerr jobs page and the DataAnnotation job board, read on 2026-11-17 and 2026-11-09, do not describe exposure to sensitive content. The court record differs: Schuster v.
Scale AI, Inc., et al. (N.D. Cal. 3:25-cv-00620-JSC) began with a complaint filed 2025-01-17 (Dkt. 1), a class action alleging psychological harm from writing violent and abusive prompts.
The plaintiffs filed a notice of voluntary dismissal dated 2026-08-10, with prejudice as to the named plaintiffs and without prejudice as to the putative class. It was a dismissal, not a settlement, and the allegations were not decided.
Press accounts of annotator pay and harm are not stated here. The court history sits with contributor lawsuits against human data platforms.
What the pages do not say about rights in a submitted attack
None of the program pages read states who owns a submitted attack, whether it may be reused in training or sold, or how long it is kept.
The Anthropic announcement and the Gray Swan homepage, read on 2026-11-17, carry no rights clause, and Gray Swan's Arena rules were not read.
The one recorded assignment clause for platform work is Outlier's Terms of Use s.4(b), effective 2026-05-11, which irrevocably assigns all rights in Work Product, and it is not stated to cover red-teaming tasks.
The questions to answer from a program's current rules before submitting are who owns the attack, whether customers may reuse it, how long it is retained, who may see the content, when payment is due and which locations are excluded.
Benefits, risks, downsides and alternatives of red-teaming work
Red-teaming work lets people with security, safety or creative-adversarial skill be paid for finding model failures, but it runs on terms each program sets, with exposure to harmful content, invite-only access and unstated rights.
Benefits
- Stated rewards and prizes: a bounty ceiling of USD 15,000 (Anthropic, announced 2024-08-08) and cash prizes (Gray Swan).
- Early access to an unreleased safety system in the Anthropic bounty.
- Work that uses adversarial skill to improve model safety.
Risks
- Exposure to harmful content, with precautions reported in one 2022 paper and an allegation in one court case.
- Rights in submissions not stated on the pages read.
- Tight eligibility and invite-only access.
- Claims that differ across a program's own pages, as with Gray Swan's two sets of figures.
Downsides
- Pay stated as a ceiling or not at all.
- Screening and application effort that is unpaid where an application is stated.
- Invite-only programs that may be closed, with the Anthropic bounty's current status not stated.
Alternatives
- Rating or expert tasks on other platforms, linked below.
- Security research outside AI.
- Employment in a safety or security team.
- Not taking the work.
How this page is compiled
Programs come from the organizations' own pages read on 2026-11-17 and the court record, and papers from the authors' arXiv postings.
The table is alphabetical and carries no ranking. Forums, job boards, listings, salary pages and press are not sources, and a page that could not be read (OpenAI, Scale AI) is shown as not read.
Where red-teaming sits among human data and who buys it
This page covers red-teaming data and work; the pages below cover the wider human data market, platforms, rating and preference data, expert fields, scams and the lab that ran the bounty.
Employer rules for side work are covered under AI training work next to a day job.
- Human data for AI: the companies: the companies behind human data
- AI training jobs and platforms compared: terms, screening and court record
- AI raters and evaluation data: rating work and how labs buy evaluation data
- Preference and ranking data for AI: ranking data, the neighbor of attack data
- Expert contributor work by field: security and domain experts by field
- Is AI training work legit: scams that imitate paid red-teaming offers and how to check a program
- Anthropic data buyer profile: the lab that ran the bounty, with its other program terms
- How DataPacts is funded: how the site is funded
Frequently asked questions
What is the purpose of AI red teaming?
AI red teaming finds and reduces harmful model outputs before and after release. The Anthropic authors' 2022 paper states that aim and releases 38,961 attacks as a dataset, and rated attacks can also serve as training material.
How much do AI red teamers get paid?
Pay is stated only as claims: Anthropic's bounty offers up to USD 15,000 per novel universal jailbreak (announced 2024-08-08), and Gray Swan states cash prizes without amounts (2026-11-17). The pay page holds the platform rates.
Can anyone join an AI red-teaming program?
Not on the pages read. Anthropic's 2024 bounty was invite-only after an application, Gray Swan's homepage states no criteria, and OpenAI's network page could not be read (HTTP 403, 2026-11-17).
Which companies run AI red teaming?
Organizations that run human programs are in the table: Anthropic with its bounty and Gray Swan with its Arena. Ranked lists of red-teaming software are a different market.
Is red-teaming work harmful to the people who do it?
The sources disagree in kind, not in size. The 2022 paper reports precautions and a favorable survey, one court docket records an allegation, and the platform pages read say nothing.