Red‑Teaming Quality Assurance Lead (Remote)

About the role

This remote, hourly contract position leads quality assurance for AI red‑teaming and safety‑evaluation projects. You will ensure that safety training data is realistic, policy‑aligned, and useful for identifying model vulnerabilities.

Key responsibilities

• Oversee quality, consistency, and trainer performance across AI red‑teaming projects.
• Review AI‑generated safety evaluations, adversarial prompts, risk analyses, and trainer/QA work.
• Evaluate output quality against detailed project guidelines and rubrics.
• Provide precise written feedback and communicate updates to trainers and QAs.
• Identify recurring quality issues, support onboarding, maintain documentation, and reactivate inconsistent contributors.
• Assess work for risk identification, adversarial reasoning, policy awareness, safety taxonomy alignment, prompt quality, scenario realism, vulnerability coverage, clarity, formatting, and instruction‑following.

Required profile

• Bachelor’s, Master’s, or professional experience in Computer Science, Cybersecurity, AI Safety, Trust & Safety, Public Policy, Psychology, Linguistics, Law, Security Studies, Risk Analysis, or a related field.
• Strong command of English for following guidelines, team communication, and written feedback.
• 3+ years of experience in AI safety, red‑teaming, cybersecurity, trust & safety, content policy, risk analysis, adversarial testing, model evaluation, or content moderation.
• Deep understanding of AI risk categories, adversarial prompting, jailbreak patterns, harmful‑content taxonomies, misuse scenarios, policy interpretation, model behavior, and safety evaluation principles.
• Ability to evaluate red‑teaming content against detailed rubrics and spot issues such as weak adversarial design, unrealistic scenarios, poor risk categorization, policy misinterpretation, unsafe outputs, or superficial vulnerability testing.

Required skills

• AI safety
• Red‑teaming
• Risk analysis
• Content moderation
• Adversarial testing
• Model evaluation