AI Safety Experts — English & Malay

$17–25/hrrate as listed

Mercor is recruiting bilingual red teamers who probe conversational AI models in English and Malay, attempting jailbreaks, prompt injections, and multi-turn manipulation, then documenting what broke and why. Observed pay for this project runs $17–25/hr, fully remote and text-based, with sensitive-topic work clearly flagged in advance and optional.

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

The work

You spend your sessions adversarially probing a model: constructing prompts designed to bypass refusals, testing whether safety behavior holds up when the same request is translated into Malay or code-switched mid-conversation, escalating across multiple turns, and looking for bias or misinformation failures that a single-shot automated test would never surface. When something breaks, the deliverable is not the anecdote — it's a reproducible artifact. You annotate the failure, classify the vulnerability against the project taxonomy, note severity and reproducibility, and write it up so a customer's safety team can act on it without asking you follow-up questions.

Structure matters more than cleverness here. Projects supply benchmarks, playbooks, and vulnerability taxonomies, and reviewers care whether your attack cases map cleanly onto them. Malay-language work is a specific reason this role exists: safety training is often thinner outside English, and failures that appear only in Malay — or in the seam between the two languages — are exactly the coverage gap the project is buying.

What the screen looks for

Verifiable prior red teaming. AI adversarial work, penetration testing, exploit development, abuse or disinformation analysis, or structured socio-technical probing. Expect to describe specific engagements and what you actually found.
Genuine native-level fluency in both English and Malay, including register, slang, and regional variation — not textbook competence.
Method, not vibes. You should be able to explain how you decide what to test next, and how you tell a real vulnerability from a lucky one-off output.
Clear writing under constraint. Reports are the product.

Logistics

Remote and asynchronous. Work arrives as project batches with defined taxonomies and deadlines; hours are flexible and vary by project demand rather than being guaranteed weekly. All content is text-based. Some batches involve sensitive material — bias, harassment, misinformation, harmful behaviors — and topics are communicated before exposure, with participation in higher-sensitivity work optional and wellness resources available.

What they look for

Prior red teaming experience (AI adversarial testing, cybersecurity, or socio-technical probing)
Native-level fluency in both English and Malay
Working knowledge of jailbreaks, prompt injection, and multi-turn manipulation techniques
Ability to apply structured taxonomies, benchmarks, and testing playbooks consistently
Clear technical writing and reproducible documentation of attack cases
Comfort reviewing sensitive content including bias, misinformation, and harmful outputs

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 both English and Malay?
Native fluency in English and Malay · required
Do you have documented experience in red teaming, adversarial machine learning testing, cybersecurity penetration testing, or similar adversarial work on real systems?
Prior red teaming or adversarial testing experience · required
Are you able to work remotely with flexible hours compatible with this role's remote-first structure?
Remote work availability · preferred

0 of 3 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 have genuine hands-on red teaming or adversarial testing experience, not just theoretical knowledge of AI safety. Mercor needs people who've already hunted for breakage.
A strong answer: Your answer must name a concrete system you probed, describe the specific vulnerability you surfaced (not 'I found a bug'), and show that you documented or reported it in a way someone else could act on. The vulnerability should involve non-obvious reasoning—not a typo.
Where people lose it: Talking about vulnerabilities you read about, attended a conference talk about, or 'would' test, rather than something you personally uncovered. Generic statements like 'I'm good at finding edge cases' without a real example.