AI Safety Experts — English & Odia

What you would do

• Red team conversational AI models by probing jailbreaks, prompt injections, and misuse cases
• Test for bias, cultural insensitivity, misinformation, and inappropriate responses
• Annotate failures, classify vulnerability types, and flag systemic risks using taxonomies
• Generate reproducible attack datasets and cases customers can learn from
• Document testing methodology and findings clearly for technical and non-technical audiences

Who they want

• Native fluency in English and Odia required for bilingual testing and cultural assessment
• Strong judgment about language accuracy, completeness, and audience appropriateness
• Rigorous attention to detail catching subtle errors, inconsistencies, and logical gaps
• Structured approach to guidelines, playbooks, and quality standards without ad hoc decisions
• Clear communication explaining reasoning and findings to diverse stakeholders

Main skills

What the interview asks about

1. 1.Jailbreak and prompt injection tactics

   Red teaming means thinking adversarially. Interview checks whether you understand exploitation techniques and can apply them to new domains.

   For example: “Describe a jailbreak attempt you've tried on a conversational AI. Why did you think it might work? How did the model respond? What did you learn?”

2. 2.Language-specific bias detection

   You're fluent in both English and Odia. Interview assesses whether you can spot culturally specific biases or language-dependent vulnerabilities.

   For example: “An AI generates a response in Odia that's grammatically correct but disrespectful toward a particular group. How would you classify this in your taxonomy? What makes it a vulnerability?”

3. 3.Subtle accuracy failures

   Your job is catching what automated tests miss: fluent language hiding logical gaps or misinformation. Interview checks rigor and discernment.

   For example: “A conversational AI answers a factual question with fluent, well-structured prose that contains one critical error buried in the middle. How would you document it for developers?”

4. 4.Reproducible case documentation

   Finding a bug is half the work. Interview checks whether you write reports that let engineers fix it quickly without needing clarification.

   For example: “You discover a jailbreak that works 70% of the time. How would you structure your documentation so the engineering team can reproduce it and prioritize the fix?”

5. 5.Multi-turn conversation exploitation

   Single-turn jailbreaks are obvious. Real attacks manipulate across multiple exchanges. Interview probes whether you understand cumulative context.

   For example: “Design a three-turn attack where the first two messages prime the model, then the third triggers a vulnerable response. Explain why the progression matters.”

A task you may get

You probe an AI's handling of a sensitive cultural topic in both English and Odia. Document one successful attack path showing how phrasing or language choice leads to an unsafe response. Classify the vulnerability and explain why it matters.

How to prepare

• Study common jailbreak techniques: role-play prompts, hypotheticals, context overrides. Practice adapting them to conversational scenarios.
• Prepare examples of culturally specific or language-dependent biases from your own experience or research.
• Review vulnerability taxonomies used in adversarial ML: know how to classify by type, severity, and exploitability.
• Write three detailed documentation examples: one successful attack, one bias finding, one misinformation case. Focus on clarity and actionability.

The facts

Pay

$16–22/hr

Hours

Hourly, 40 hours a week

Where

Remote

Field

Data Analysis

Project name

Neon

Posted

6/4/2026

Places left

15

We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.