AI Safety Experts — English & Danish

$48–62/hrrate as listed

Mercor is staffing a bilingual red team to probe conversational AI models in English and Danish, documenting jailbreaks, prompt injections, bias exploits, and multi-turn manipulation. Work is remote, text-based, and paid in an observed range of $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.

What the work involves

You attack models on purpose. A typical session means picking a threat category from a provided taxonomy — jailbreaks, indirect prompt injection, misuse elicitation, bias exploitation, multi-turn social engineering — and running structured adversarial conversations against a target model or agent in English, Danish, or across the two. Danish matters here because safety guardrails are usually tuned on English data; refusals, slang, euphemism, and cultural framing often behave differently once you switch languages or code-switch mid-conversation. When something breaks, you annotate the failure, classify the vulnerability, and write it up so someone else can reproduce it from your notes alone.

The output is data, not vibes. Attack cases, labeled transcripts, severity judgments, and short reports go to customers who use them to patch systems. Consistency is graded as heavily as creativity: two experts working the same taxonomy should classify the same failure the same way.

What the screen looks for

Verifiable red teaming history — adversarial ML work, penetration testing, abuse/trust-and-safety analysis, or socio-technical probing. Expect follow-ups asking what you actually did, not what your team did.
Native-level Danish and English, with the ability to construct adversarial prompts that exploit linguistic and cultural nuance rather than translating English attacks word-for-word.
Method over improvisation — can you describe a framework, benchmark, or coverage strategy you've used, and explain why a random-hack approach undertests a model?
Judgment on severity and reproducibility — distinguishing a genuine safety failure from a one-off sampling artifact, and knowing when a finding is worth escalating.

Logistics and content exposure

Fully remote and asynchronous; you take batches against deadlines rather than sitting in a shift. Contributors commonly work 10–25 hours per week, though volume fluctuates with customer demand and no minimum is guaranteed. Pay in the $48–62/hr band is observed on comparable Mercor safety projects and varies by specialization and task type.

The work involves reading and generating content touching bias, misinformation, harassment, and other harmful behaviors. All content is text-based. Higher-sensitivity streams are opt-in, topics are disclosed before exposure, and wellness resources are provided — declining a sensitive stream does not disqualify you from the project.

What they look for

Native fluency in both Danish and English
Prior red teaming experience (adversarial ML, penetration testing, or socio-technical probing)
Working knowledge of LLM attack surfaces: jailbreaks, prompt injection, multi-turn manipulation
Structured testing against taxonomies, benchmarks, or coverage playbooks
Clear written documentation of reproducible attack cases and severity rationale
Trust-and-safety, abuse analysis, or content-policy background (preferred)

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 English speaker, or do you have native-level fluency in English?
Native English fluency · required
Are you a native Danish speaker, or do you have native-level fluency in Danish?
Native Danish fluency · required
Do you have documented hands-on experience in at least one of the following: AI red teaming, adversarial ML testing, cybersecurity penetration testing, or socio-technical vulnerability probing?
Prior red teaming or adversarial work · required
Are you able to work remotely and maintain consistent availability for a text-based role?
Remote work eligibility · preferred
Are you comfortable working with AI outputs on sensitive topics such as bias, misinformation, or harmful behaviors, with the understanding that participation in higher-sensitivity projects is optional?
Comfort with sensitive content · preferred

0 of 5 answered

Included with this listing

Interview prep — 10 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: Tests whether the candidate has *hands-on* adversarial experience (not just theory) and whether they can articulate a structured attack methodology rather than random fuzzing.
A strong answer: The candidate describes a concrete incident where they probed a system methodically (e.g., targeting a specific failure mode, using a taxonomy or framework to guide the attack, iterating on inputs), identified a real gap, and explained why it mattered — not a vague claim of 'finding bugs.'
Where people lose it: Conflating security awareness with red team work — saying 'I know AI can be jailbroken' or 'I've read about prompt injection' without ever having *actually attempted* to craft or test an attack.