AI Safety Red Teamer

$70–84/hrrate as listed

Design adversarial prompts that probe frontier models for jailbreaks, unsafe outputs, and policy failures across sensitive domains like cyber, biosecurity, and misinformation. Mercor lists this at $70–84/hr, contract and remote, with work documented as reproducible vulnerability write-ups rather than one-off prompt attempts.

Summary, eligibility guidance and screening preparation are prepared by Sidequest from the platform's posting. Check the source listing before applying.

What the work actually involves

You work against a model behind a testing interface, usually with a policy document and a target domain assigned for the session. The job is to find the boundary — where a model refuses correctly, where it refuses too much, and where it complies with something it shouldn't. Most attempts fail, which is the point; the deliverable is the small share that succeed, documented well enough that someone else can reproduce the failure and a trainer can build a fix from it.

A typical write-up includes the attack framing, the full transcript, why the output violates policy, how many attempts it took, and whether variations of the prompt still work. Multi-turn escalation, role-play framings, encoding tricks, hypothetical wrappers, and legitimate-professional-context framings are all standard tools. Depth in one grey-area domain matters more than breadth: someone who knows what actually constitutes uplift in biosecurity or what a working exploit chain looks like writes findings that researchers can act on.

What the screen looks for

Verifiable background. Five-plus years in AI safety, trust & safety, cybersecurity, investigative journalism, or a life science — with specifics about what you handled and at what level of severity.
Domain depth under follow-up. Expect the interview to push past your first answer on your claimed specialty; general familiarity reads as thin quickly.
Calibration. Whether you can distinguish a genuinely harmful output from something merely uncomfortable, and whether you recognize over-refusal as a failure mode too.
Written clarity. Red teaming output is a document. Vague or dramatized reports have little value.

Logistics

Remote and asynchronous, contract-based, with hours logged against assigned campaigns. Most contributors run 10–25 hours weekly; volume fluctuates as campaigns open and close. Work is confidential, and NDA plus attestation to handling sensitive material is standard. Content is deliberately unpleasant in places — misinformation, fraud, self-harm, CBRN adjacency — and the platform expects you to be honest with yourself about that before applying.

What they look for

5+ years in AI safety, trust & safety, cybersecurity, investigative journalism, or a life science
Demonstrated experience designing adversarial prompts or jailbreak tests against frontier models
Deep expertise in at least one grey-area domain (cyber, biosecurity, misinformation, political content, fraud)
Clear technical writing — reproducible vulnerability documentation
Bachelor's degree or higher in a relevant discipline
Familiarity with model policy frameworks, RLHF/SFT, or adversarial evaluation methodology

Before you apply

Would you pass this screen?

Drawn from what this listing actually requires. Answered in your browser — nothing is sent to us, and we do not store your answers.

Do you hold a Bachelor's degree or higher in Computer Science, Cybersecurity, Journalism, Communications, Psychology, Biology, Chemistry, Public Policy, or a related discipline?
Bachelor's degree or equivalent · required
Do you have 5 or more years of professional experience in AI Safety, AI Red Teaming, Trust & Safety, cybersecurity, investigative journalism, life sciences, or a closely related field?
5+ years professional experience in relevant field · required
Have you designed adversarial prompts or performed hands-on evaluation of frontier AI systems as part of a professional role or project?
Adversarial prompt design or AI system evaluation experience · required
Do you have substantive knowledge in at least one grey-area domain such as cybersecurity, biosecurity, misinformation, political content, fraud, or scientific safety?
Grey-area domain expertise · preferred

0 of 4 answered

Included with this listing

Interview prep — 11 questions

Matched to what this platform's screening actually measures. Answer each out loud in 60–90 seconds: claim → concrete example → implication.

What it tests: Whether the candidate has genuine adversarial testing experience and can articulate the methodology behind finding flaws, not just listing them. Red teaming requires structured thinking about where systems fail.
A strong answer: The answer should describe a specific vulnerability discovery (in AI, security, content moderation, or a closely related domain), explain the deliberate approach or reasoning that led to finding it, and show awareness of why that flaw mattered—not a lucky find, but a methodical one.
Where people lose it: Candidates often describe finding bugs through normal use or passive observation, rather than active adversarial design. Red teaming is not quality assurance; confusing the two signals a fundamental misunderstanding of the role.