AI Safety Experts — English & Dutch

$48–62/hrrate as listed

Mercor is staffing a bilingual red team to probe conversational AI models in English and Dutch, surfacing jailbreaks, prompt injections, bias exploits, and multi-turn manipulation. Work is text-based and remote, with observed rates in the $48–62/hr range depending on prior red teaming depth.

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

The work

You attack models on purpose. A typical session means picking a threat category from a provided taxonomy — misuse, disinformation, harassment, bias exploitation, agentic misbehavior — and building attack chains against a conversational model or agent until something breaks or you can credibly show it doesn't. Dutch matters here because safety behavior degrades across languages: a refusal that holds firmly in English often collapses in Dutch phrasing, idiom, or code-switching, and part of your job is finding exactly where that asymmetry lives.

When you find a failure, the deliverable is not the screenshot. It's a reproducible artifact: the prompt sequence, the conditions under which it fires, a classification against the project's vulnerability schema, a severity judgment, and enough annotation that an engineer who has never seen your conversation can replicate it. Reproducibility is the metric that separates useful red team data from anecdote.

Some projects touch sensitive content — bias, self-harm adjacent material, harmful instruction requests, misinformation. Mercor states that topics are communicated before exposure, higher-sensitivity work is opt-in, and wellness resources are available. Treat that as a real choice rather than a formality.

What the screen looks for

Genuine native fluency in both languages, not conversational competence. Expect to be probed on Dutch register, regionalisms, and how you'd construct an attack that only works in Dutch.
Prior adversarial experience with specifics: penetration testing, adversarial ML research, trust & safety investigation, disinformation analysis, or documented AI red teaming.
Structured method over improvisation. Screens tend to reward candidates who reference frameworks, coverage thinking, and systematic variation over people who describe one clever jailbreak.
Clear risk communication — can you explain severity to someone non-technical without inflating it.

Logistics

Fully remote and largely asynchronous. Work is contract and project-based, meaning volume fluctuates with customer demand; several contributors run this alongside other commitments at 10–25 hours per week, though full-time loads appear during active campaigns. Rates in the $48–62/hr band are what has been observed for this posting and vary by specialty depth and project — nothing here is guaranteed.

What they look for

Native-level fluency in both English and Dutch
Prior red teaming experience — AI adversarial testing, penetration testing, or socio-technical probing
Ability to work systematically from taxonomies, benchmarks, and structured playbooks
Clear written documentation of reproducible attack cases and vulnerability classifications
Familiarity with LLM failure modes: jailbreaks, prompt injection, multi-turn manipulation, bias exploitation
Comfort reviewing sensitive content including harmful, biased, or misleading material

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.

Are you a native speaker of English?
Native English fluency · required
Are you a native speaker of Dutch?
Native Dutch fluency · required
Do you have hands-on experience in red teaming, adversarial testing, jailbreaking, or penetration testing—in AI, cybersecurity, or another domain?
Prior red team or adversarial testing experience · required
Are you able to work reliably in a timezone compatible with remote collaboration during business hours?
Remote work timezone compatibility · 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 red team or adversarial testing experience, and whether they apply structure (frameworks, taxonomies, reproducibility) rather than ad-hoc probing.
A strong answer: A strong answer names a specific vulnerability discovery (in AI, security, or a complex system), explains the structured methodology used to surface it (e.g., a testing framework, threat model, or taxonomy), and shows how findings were documented or acted upon by stakeholders.
Where people lose it: Candidates claim red teaming experience but describe only bug reports, customer complaints, or general 'testing'—not active adversarial probing with intent to break the system. Or they describe finding a flaw but cannot explain the repeatable process that led to it.