Risk Analyst III

Job Title:

Risk Analyst III Duration:

12 Months - chances of extension

Location:

Open to any office in EST but highly preferred is Austin, TX. Overview About the Role We are seeking creative, resilient, and highly motivated AI Red Teamers to join our Red Teaming team. In this role, you will be at the forefront of AI safety, identifying and mitigating risks in our advanced language models. Unlike automated testing, which handles baseline coverage and known attack patterns efficiently, this role focuses on the ~40% of vulnerabilities that require human ingenuity, psychological insight, and creative out-of-distribution thinking. You will interact with our models across text, image, audio, and video modalities to uncover weaknesses, evaluate actual output harm, and stress-test our systems against novel adversarial attacks. This is intense, high-impact work at the frontier of AI safety — and it offers direct influence on how Client's AI systems behave in the real world. Important Notice This role involves exposure to graphic and/or objectionable content, including but not limited to graphic images, videos, audio, and writings; offensive or derogatory language; and other potentially disturbing material such as child exploitation, graphic violence, self-injury, and animal abuse. Testing may also require verbalizing model prompts containing references to such content. Wellness infrastructure and opt-out policies are in place to support all team members. What You'll Do Creative Adversarial Testing Design and execute novel, multi-turn adversarial attacks — including emotional manipulation, roleplay, social engineering, and authority exploitation — to bypass model safeguards and surface harmful capabilities. Vulnerability Assessment Evaluate model outputs for actual harm and real-world risk, not just policy violations. Apply Client's user risk taxonomy to prioritize testing across user types, from casual users to agentic systems. Agentic and Emerging Threat Testing Probe for agentic vulnerabilities such as privilege escalation, indirect prompt injection, and scope creep in multi-authority systems — the next frontier of AI risk. Data Annotation and Reporting Generate high-quality human evaluation data by annotating model failures, classifying vulnerabilities, and producing reproducible adversarial test cases that engineering and safety teams can act upon. Cross-Functional Collaboration Partner with AI researchers, engineers, and domain experts to translate findings into actionable improvements. Contribute to refining our red teaming taxonomy, benchmarks, and tooling infrastructure. Continuous Learning Stay current with evolving adversarial techniques, internet subcultures, and AI safety research to continuously sharpen your attack strategies. What We're Looking For We actively seek candidates from non-traditional backgrounds. Our most effective red teamers come from creative, humanities, mental health, and special education fields — not exclusively from technical roles.

Minimum Qualifications Creative and Psychological Insight:

Background in creative writing, humanities, mental health counseling, psychology, or special education — with a demonstrated ability to construct compelling narratives, exploit linguistic nuance, and identify psychological vulnerabilities.

Adversarial Mindset:

A natural inclination to think like an attacker and push systems to their limits. You should find genuine satisfaction in discovering unexpected failure modes.

Adaptability:

Comfort switching between modalities (text, image, audio, video) and rapidly adjusting to new model behaviors, testing priorities, and task types.

Communication Skills:

Excellent written and verbal communication skills, with the ability to document and explain complex vulnerabilities clearly to both technical and non-technical audiences.

Resilience and Balance:

Capacity to sustain well-being while engaging in psychologically demanding work. This role involves regular exposure to graphic and objectionable content, including violence, exploitation, and self-harm scenarios. Comprehensive wellness support is provided. Preferred Qualifications Prior experience in professional red teaming, trust & safety, data annotation, or socio-technical risk analysis. Familiarity with large language models (LLMs) and generative AI products such as ChatGPT, Claude, or Gemini. Basic technical skills in prompt engineering, encoding techniques (e.g., Base64, ROT13), or scripting to complement creative attack vectors. Knowledge of AI safety concepts, including RLHF, alignment, and model evaluation frameworks.