AI Safety Experts — English & Indonesian

$17–25/hrrate as listed

Red team conversational AI models in English and Indonesian, probing for jailbreaks, prompt injections, bias exploitation, and multi-turn manipulation. Work is text-based, contract, and remote, with observed rates of $17–25/hr on Mercor.

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

The work

You are the adversary. Day to day, that means opening a conversation with a model and methodically trying to make it fail — direct jailbreak attempts, indirect prompt injection, role-play framing, incremental multi-turn escalation, and language-switching attacks that exploit weaker safety coverage in Indonesian than in English. Cross-lingual probing is a large part of why this cohort exists: safety training that holds in English often degrades in Bahasa Indonesia, in code-switched prompts, or in regionally specific slang, idiom, and cultural framing. You will be asked to find exactly those seams.

Every successful attack becomes an artifact. You annotate what failed and why, classify the vulnerability against a provided taxonomy, and write the case up so an engineer who has never seen your conversation can reproduce it. Consistency matters more than cleverness here — the project runs on shared benchmarks and playbooks, and an attack that cannot be replicated is not deliverable data. Expect regular guideline updates, calibration rounds, and reviewer feedback on your write-ups.

What the screen looks for

Mercor's process is largely AI-led: a resume parse, then a structured voice or chat interview, then task-based calibration. The interview probes for verifiable red teaming history — adversarial ML work, penetration testing, trust and safety, abuse analysis, or socio-technical research — and pushes on specifics. Vague claims about "testing chatbots" collapse under follow-up; naming a taxonomy you used, an attack class you specialize in, or a failure you documented does not. Genuine native-level fluency in both English and Indonesian is verified in conversation, not self-reported.

Logistics and content exposure

Fully remote and asynchronous; you claim tasks against open batches rather than working fixed shifts.
Hours are variable and project-driven — some contributors work 10–15 hours weekly, others more when a customer engagement ramps.
Observed pay is $17–25/hr, positioned by experience and language coverage. Rates are as reported, not guaranteed.
The work involves reading and generating text touching bias, misinformation, and harmful behaviors. Topics are disclosed before exposure, higher-sensitivity streams are opt-in, and wellness resources are provided.

What they look for

Native-level fluency in both English and Indonesian, including code-switched and colloquial registers
Prior red teaming experience — AI adversarial testing, penetration testing, or socio-technical probing
Working knowledge of LLM attack classes: jailbreaks, prompt injection, multi-turn manipulation, bias exploitation
Structured documentation: reproducible attack cases, vulnerability classification against a taxonomy
Ability to follow and apply evaluation rubrics and benchmarks consistently across large task batches
Nice to have: adversarial ML research, trust and safety, abuse analysis, or exploit development background

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 or bilingual English speaker with professional fluency?
Native English fluency · required
Are you a native or bilingual Indonesian speaker with professional fluency?
Native Indonesian fluency · required
Do you have prior experience in red teaming, adversarial testing, cybersecurity testing, or similar vulnerability assessment work?
Prior adversarial testing or red teaming experience · required
Are you able to work remotely with reliable internet access, and available to work within overlap hours with Mercor's team?
Remote access and availability · required

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, not just theoretical knowledge; whether they understand the difference between finding a bug and producing an artifact that someone else can act on.
A strong answer: The answer names a specific system (AI, security, product, or process), describes a concrete method used to uncover the flaw (not luck), and explains how the finding was documented or reproduced well enough that the owner could fix it or defend against it.
Where people lose it: Describing a vulnerability they *read about* rather than one they personally discovered, or finding something but never confirming it was real or reproducible—showing they can poke systems but not produce evidence others trust.