AI red teaming jobs: what they involve and what they pay

What AI red teaming jobs are

AI red teaming is paid, authorised testing of a model's safety behaviour before and after release. The client wants to know where its safeguards fail, so it hires people to probe them under controlled conditions and write up what they find. It is a specialist corner of evaluation work; the broader field is covered in what AI training work actually is.

On this site the market is almost entirely one platform. Of the 699 listings live on 29 September 2026, 56 mention red teaming in the title or skills. Mercor posted 55 of them and Invisible Technologies one.

What one task involves

Across the listings, the work is organised as a test campaign, and a task has three parts:

1. Write a test request within the project's scope and label what kind it is. Mercor's domain panels ask for single-turn prompts classed as benign, dual-use or adversarial.
2. Judge the response against the client's written policy: did the model help where it should, and decline where it should?
3. Write it up. Classify any failure against the project taxonomy, note whether it looks isolated or systematic, and explain it clearly enough for engineers to reproduce and fix.

Over-refusal counts as a failure too. Mercor's radiological safety panel states that refusing an ordinary professional question is as much a failure as helping with harm, so much of the effort goes into judging where the line sits.

Three kinds of listing

Bilingual testers. Mercor's "AI Safety Experts: English and X" family (18 listings) tests whether safeguards that hold in English also hold in other languages. The case is that a native speaker knows how people actually phrase things. Mercor's English and Danish listing is typical.

Generalist red teamers. Mercor's AI Safety Red Teamer asks for five or more years in AI safety, trust and safety, cybersecurity, investigative journalism, life sciences or a related field.

Domain panels. Twenty listings recruit credentialled specialists from regulated technical fields (nuclear safeguards, radiation protection, chemistry and similar) to check whether a model can tell a professional's routine question from a harmful one. These ask for responsibility in the field, such as licensing, custody or inspection experience, not just a relevant degree.

What it pays

The 32 hourly listings run from $16 to $175 an hour, and the pay tracks the tier:

• Bilingual testers: $16–62 an hour, priced by language. South Asian languages sit at $16–22; Danish, Dutch, Finnish, Norwegian and Swedish at $48–62.
• Mercor's generalist red teamer: $70–84 an hour.
• Domain panels: $65–75 per task, on all 24. None states how long a task takes, so the hourly equivalent is unknown.
• The top figure, $125–175, is a cybersecurity listing for paid expert interviews that names red teaming as one skill among several.

Mercor's listings state weekly payment via Stripe or Wise, and that it cannot support H-1B or STEM OPT candidates.

Before you apply

The material is uncomfortable by design, and the listings we read do not mention wellbeing support. The generalist listing says projects may be extended, shortened or ended early depending on needs and performance, and the work is independent contractor work on your own schedule.

Browse the current set under AI Safety & Evaluation; the bilingual listings also appear under Languages & Linguistics.