AI Safety Experts — English & Swedish

$48–62/hrrate as listed

Mercor is staffing a bilingual red team to probe conversational AI models in English and Swedish, documenting jailbreaks, prompt injections, and multi-turn manipulation attempts. Work is remote, text-based, and paid at an observed $48–62/hr, with sensitive-content projects offered on an opt-in basis.

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 will run adversarial sessions against conversational models and agents, working in both English and Swedish. A typical batch might ask you to attempt a category of jailbreak across ten seeded scenarios, escalate across multiple turns, and record where the model's refusal behavior degrades. Cross-lingual work matters here: attacks that fail in English often succeed in Swedish, and part of the value you provide is documenting that asymmetry — idiom, code-switching, dialect, and translation-layer gaps included.

Every successful probe becomes an artifact. You classify the failure against a provided taxonomy, annotate severity, and write a reproduction path clear enough that an engineer who does not speak Swedish can verify it. Random clever hacks are less useful than structured coverage; reviewers weight consistency with the playbook alongside novelty.

Multi-turn manipulation, prompt injection, misuse elicitation, bias exploitation
Failure annotation and vulnerability classification against fixed taxonomies
Reproducible write-ups: attack case, transcript, severity rationale
Flagging systemic patterns rather than one-off outputs

What the platform screens for

Mercor's screen is AI-led and expects verifiable specifics. Expect follow-ups on any red teaming, security, trust-and-safety, or adversarial research you claim — what the scope was, what you found, how you documented it. Native-level Swedish is a hard requirement and is probed conversationally, not self-reported. Candidates from cybersecurity, content moderation, disinformation research, linguistics, psychology, and investigative writing all clear this bar; what they share is the habit of working from a framework and explaining risk plainly to non-specialists.

Logistics

Fully remote and asynchronous, with work delivered in batches. Contributors commonly report 10–25 hours per week, though volume moves with customer demand. Pay observed in this band is $48–62/hr; rates are set per project and are not guaranteed. Higher-sensitivity content streams — harassment, self-harm, extremism, disinformation — are opt-in, disclosed by topic before you see anything, and supported by written guidelines and wellness resources.

What they look for

Native fluency in both English and Swedish
Prior red teaming experience: AI adversarial testing, penetration testing, or socio-technical probing
Structured testing against taxonomies, benchmarks, or documented playbooks
Clear technical writing — reproducible attack cases and severity rationales
Familiarity with LLM failure modes: jailbreaks, prompt injection, multi-turn manipulation
Comfort reviewing sensitive or harmful 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 of English or have you lived and worked exclusively in English-speaking environments for 10+ years?
Native English fluency · required
Are you a native speaker of Swedish or have you lived and worked exclusively in Swedish-speaking environments for 10+ years?
Native Swedish fluency · required
Do you have documented professional experience in red teaming, adversarial AI testing, penetration testing, or security research?
Prior red teaming or adversarial testing experience · required
Are you able to work remotely and have access to a stable internet connection and private workspace?
Remote work capability · required
Are you able to commit to at least 20 hours per week for this role?
Availability for assignment · preferred

0 of 5 answered

Included with this listing

Interview prep — 13 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: Red teaming requires both adversarial creativity and rigor; this tests whether the candidate has hands-on experience finding real vulnerabilities, not just theoretical knowledge, and whether they think reproducibility matters.
A strong answer: The answer must name a specific system or context (AI model, security tool, application), describe a concrete attack or failure mode the candidate discovered, and explain how they captured or documented it in a way that could be acted on by an engineer or customer.
Where people lose it: Candidates often stay vague ('I found security issues') or describe a generic attack pattern they read about rather than something they personally executed and verified. Without reproducibility details, they signal they may not understand why Mercor needs artifacts that customers can actually use.