AI Safety Experts — English & Assamese

$20–22/hrrate as listed

Red team conversational AI models in English and Assamese, probing for jailbreaks, prompt injections, bias exploitation, and multi-turn manipulation. Work is remote, text-based, and paid at an observed $20–22/hr on Mercor.

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 spend your sessions adversarially probing conversational models and agents, then writing up what broke. A typical task starts from a taxonomy or playbook — a category of harm, a misuse pattern, an attack class — and asks you to construct prompts in English or Assamese that push the model past its guardrails. When a model fails, you annotate the failure, classify the vulnerability, and produce a reproducible case: the exact prompts, the turn sequence, the conditions under which it recurs. Multi-turn manipulation matters here; single-shot jailbreaks are the easy part, and much of the value comes from attacks that build slowly across a conversation or exploit the gap between a model's English and Assamese safety behavior.

The Assamese requirement is substantive, not cosmetic. Safety training is unevenly distributed across languages, and low-resource-language red teaming surfaces failures that English testing never reaches — code-switching, transliteration, register shifts, culturally specific framings of harm. You need native fluency in both languages to know when an output is subtly wrong rather than obviously wrong.

What the platform screens for

Mercor's screening is AI-led and interview-based, and it presses on specifics. Expect to be asked to describe an actual attack you constructed and why it worked at a mechanistic level, not just that it did. Vague claims of "red teaming experience" don't survive follow-up questions. Reviewers also look for structured thinking — whether you work from frameworks, taxonomies, or benchmarks rather than improvising — and for writing quality, since your deliverable is documentation another person has to act on. Verified fluency in Assamese is confirmed during the process.

Logistics and content exposure

Fully remote and asynchronous; you claim and complete tasks on your own schedule.
Observed pay band is $20–22/hr, stated as reported by contributors — rates vary by project and are not guaranteed.
Hours are typically flexible and project-dependent, with volume rising and falling as customer engagements start and end.
The work involves reviewing outputs touching bias, misinformation, and harmful behaviors. Topics are disclosed before exposure, higher-sensitivity projects are opt-in, and wellness resources are provided.

What they look for

Native fluency in both English and Assamese
Prior red teaming experience — AI adversarial work, penetration testing, or socio-technical probing
Working knowledge of LLM attack classes: jailbreaks, prompt injection, multi-turn manipulation, bias exploitation
Clear technical writing and reproducible documentation of findings
Ability to work from taxonomies, benchmarks, and structured testing playbooks
Nice to have: adversarial ML background (RLHF/DPO attacks, model extraction) or abuse/disinformation analysis

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 and Assamese, or native in one and fluently bilingual in the other, with recent substantive use in both languages?
Native fluency in English and Assamese · required
Have you conducted hands-on red teaming, penetration testing, jailbreak research, or systematic adversarial probing of AI systems or software—beyond using a product as an end user?
Prior red teaming or adversarial AI experience · required
Do you have a reliable internet connection and a setup capable of working remote hours, and are you available for this engagement?
Remote work setup and availability · required
Are you able to review and generate text on topics such as bias, misinformation, and harmful behaviors in a red teaming context, with support from wellness resources?
Ability to engage with sensitive content · preferred
Can you write clear technical reports and dataset documentation that explain findings and vulnerabilities in a way non-specialists can understand?
Written communication and documentation skill · preferred

0 of 5 answered

Included with this listing

Interview prep — 12 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 hands-on red teaming or adversarial work history, not just theoretical interest; ability to identify gaps that automation misses.
A strong answer: A concrete example naming the system, the approach taken, the specific vulnerability surfaced, and evidence it was novel or missed by standard defenses—not a generic 'I found a bug' story.
Where people lose it: Describing a known vulnerability from a public database or a bug found through normal QA rather than active adversarial probing; or staying vague about what the vulnerability actually was and why it mattered.