Probe conversational AI models in Marathi and English for jailbreaks, prompt injections, bias exploitation, and multi-turn manipulation, then document each failure as a reproducible attack case. Work is remote, text-based, and paid in an observed band of $20–22/hr.

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

What the work actually looks like

You are assigned a model or agent and a taxonomy of harms, and your job is to break it — in Marathi, in English, and often by moving between the two. A typical session means constructing multi-turn conversations that walk a model from a benign opening into unsafe territory, testing whether refusals that hold in English collapse when the same request is phrased in Marathi, in transliterated Marathi, or in code-mixed Marathi-English. When something fails, you write it up: the exact prompts, the turn where the guardrail gave way, the category of vulnerability, and enough detail that someone else can reproduce it on a fresh session.

The content is sensitive by design. Tasks touch bias, misinformation, harassment, disinfo, and other harmful behaviours. Mercor states that topics are communicated before exposure, that higher-sensitivity projects are opt-in, and that wellness resources are available. Take that seriously when deciding whether the work suits you.

What the screen is looking for

Genuine native-level Marathi and English — not conversational Marathi. Screens probe register, regional variation, transliteration habits, and whether you can produce adversarial phrasing that reads as natural to a native speaker.
Prior adversarial experience — AI red teaming, penetration testing, trust-and-safety abuse analysis, or socio-technical probing. The listing asks for it explicitly.
Structured method over clever one-offs. Expect questions about how you achieve coverage across a taxonomy rather than how you found one funny jailbreak.
Documentation discipline. The deliverable is a reproducible artifact, not an anecdote.

Logistics

Fully remote and asynchronous, organised as project-based work across different customers. Hours are typically flexible and self-scheduled, with volume varying by project; expect to state a realistic weekly commitment and to be held to it. Pay has been observed at $20–22/hr and is not guaranteed — it varies by project, and assignments can shift as customers rotate.

AI Safety Experts — English & Marathi

What the work actually looks like

What the screen is looking for

Logistics

Interview prep — 10 questions