AI Safety Experts — English & Bengali

Neon; United States; Remote; USD 120,000 - 180,000; Full time

Job summary

Mercor is assembling a red team to probe AI models with adversarial inputs. This remote role focuses on reviewing AI outputs for bias, misinformation, and harmful behavior, with clear guidelines and wellness resources provided.

You'll generate high-quality human data, annotate failures, and document reproducible reports and datasets to strengthen customers' AI systems across multiple projects and audiences.

Qualifications

• Strong judgment about language and content; can judge AI responses for accuracy and appropriateness.
• Rigorous; notices subtle errors, inconsistencies, and gaps.
• Structured; works to guidelines and quality standards.
• Communicative; explains reasoning clearly to technical and non-technical audiences.
• Adaptable; able to move across projects and customers.

Responsibilities

• Red team conversational AI models to identify jailbreaks, prompt injections, misuse cases, and bias.
• Generate high-quality human data: annotate failures, classify vulnerabilities, and flag systemic risks.
• Apply structure: follow taxonomies, benchmarks, and playbooks for consistent testing.
• Document reproducibly: produce reports, datasets, and attack cases for customers.

Skills

Adversarial ML

Cybersecurity

Socio-technical risk

Creative probing

Language judgment

Structured workflow

Job description

Location: Remote

Fluent Language Skills Required: English & Bengali. Native fluency in English and Bengali is required for this position.

Why This Role Exists

At Mercor, we believe the safest AI is the one that’s already been attacked — by us. We are assembling a red team for this project - human data experts who probe AI models with adversarial inputs, surface vulnerabilities, and generate the red team data that makes AI safer for our customers.

This project involves reviewing AI outputs that touch on sensitive topics such as bias, misinformation, or harmful behaviors. All work is text-based, and participation in higher-sensitivity projects is optional and supported by clear guidelines and wellness resources. Before being exposed to any content, the topics will be clearly communicated.

What You’ll Do

• Red team conversational AI models and agents: jailbreaks, prompt injections, misuse cases, bias exploitation, multi-turn manipulation

• Generate high-quality human data: annotate failures, classify vulnerabilities, and flag systemic risks

• Apply structure: follow taxonomies, benchmarks, and playbooks to keep testing consistent

• Document reproducibly: produce reports, datasets, and attack cases customers can act on

Who You Are

• You bring strong judgment about language and content: you can tell whether an AI response is accurate, complete, and appropriate, and explain why

• You’re rigorous: you notice subtle errors, inconsistencies, and gaps that others skim past

• You’re structured: you work to guidelines and quality standards consistently, not ad hoc

• You’re communicative: you explain your reasoning clearly to technical and non-technical audiences

• You’re adaptable: you thrive moving across projects, task types, and customers

Nice-to-Have Specialties

• Adversarial ML: jailbreak datasets, prompt injection, RLHF/DPO attacks, model extraction

• Cybersecurity: penetration testing, exploit development, reverse engineering

• Socio-technical risk: harassment/disinfo probing, abuse analysis, conversational AI testing

• Creative probing: psychology, acting, writing for unconventional adversarial thinking

What Success Looks Like

• You uncover vulnerabilities automated tests miss

• You deliver reproducible artifacts that strengthen customer AI systems

• Evaluation coverage expands: more scenarios tested, fewer surprises in production

• Mercor customers trust the safety of their AI because you’ve already probed it like an adversary

Why Join Mercor

• Build experience in human data-driven AI red teaming at the frontier of safety

• Play a direct role in making AI systems more robust, safe, and trustworthy