AI Safety Experts — English & Finnish

$48–62/hrrate as listed

Mercor is staffing a bilingual red team to probe conversational AI models in English and Finnish, documenting jailbreaks, prompt injections, and multi-turn manipulation attempts. Work is text-based, remote, and paid at an observed $48–62/hr depending on background and project.

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 target behavior — a policy the model is supposed to hold, a refusal it should maintain, a factual claim it should not fabricate — and then working through structured attempts to break it: role-play framing, prompt injection through pasted content, incremental multi-turn escalation, translation and code-switching between English and Finnish, and exploitation of bias or cultural blind spots. When something breaks, you write it up so someone else can reproduce it: the exact prompts, the turn sequence, the classification against the project taxonomy, and an assessment of severity.

The Finnish requirement is load-bearing. A great deal of safety tuning is done in English, and models frequently behave differently once a request crosses into a lower-resource language. You will be probing exactly that gap — testing whether Finnish phrasing, idiom, morphology, or culturally specific framing routes around guardrails that hold in English. Native fluency in both languages is what makes the comparison meaningful.

What the screen looks for

Prior adversarial work you can describe concretely. AI red teaming, penetration testing, exploit development, trust-and-safety abuse analysis, or disinformation research. Expect follow-ups asking what you actually tried and what failed.
Method over improvisation. Interesting one-off jailbreaks matter less than being able to explain your coverage strategy and why you tested what you tested.
Documentation discipline. Reproducibility is the deliverable. Vague findings have no downstream value.
Genuine bilingual depth, including register, slang, and regional variation — not textbook Finnish.

Logistics and content

Remote and asynchronous, with contributors typically committing 10–25 hours per week against project deadlines rather than fixed shifts. Everything is text-based. The material touches bias, misinformation, harassment, and harmful behaviors; higher-sensitivity tracks are opt-in, topics are disclosed before exposure, and wellness resources are provided. The stated band is $48–62/hr as observed on the platform — actual rates vary by specialty, project, and assessment outcome, and are not guaranteed.

What they look for

Native fluency in both English and Finnish
Prior red teaming experience — AI adversarial testing, cybersecurity, or socio-technical probing
Structured testing against taxonomies, benchmarks, or threat models
Clear written documentation of reproducible attack cases
Familiarity with LLM failure modes: jailbreaks, prompt injection, multi-turn manipulation
Comfort working with sensitive content under defined guidelines

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 or fully fluent in both English and Finnish, able to work professionally in both languages?
Native fluency in English and Finnish · required
Have you conducted red teaming, adversarial testing, penetration testing, or similar security/AI probing work before?
Prior red teaming or adversarial AI experience · required
Are you able to work fully remote and do you have reliable internet access and a suitable work environment?
Remote work access · required
Are you willing to engage with testing scenarios involving sensitive topics such as bias, misinformation, or harmful content, with clear guidelines and wellness support?
Willingness to handle sensitive content · 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 you've done adversarial work before (AI jailbreaks, pentesting, or socio-technical probing) and whether you think in terms of *reproducible, documented findings* rather than just
A strong answer: Name a specific system, describe a concrete failure you surfaced (jailbreak, prompt injection, bias exploitation, or security flaw), and explain how you documented or reported it in a way the owning team could act on—ideally with metrics or a clear reproduction path.
Where people lose it: Vague storytelling ('I found a bug') with no technical specificity, no mention of how you documented it, or conflating finding a flaw with just pointing it out without evidence.