СШАAI Red Teamer (LLM Generalist)
1 мес. назад
32–95 USD / часСШАOnsite
prompt engineering
Stress-test large language models using adversarial prompts to expose vulnerabilities and support AI safety.
О компании
- AI partners with leading AI research labs to make models safer and more robust. Our red teaming operations help identify vulnerabilities before they reach users, contributing directly to the responsible development of frontier AI systems.
Обязанности
- As an AI Red Teamer, you will stress-test large language models by intentionally trying to break them. Rather than checking whether an answer is correct, you will design creative, adversarial prompts that expose vulnerabilities: unsafe content, bias, broken guardrails, hallucinations, prompt injection weaknesses, and unexpected behaviors. Your work directly supports AI safety and model robustness for leading research labs.
- This is a generalist red teaming role. You will probe models across the full spectrum of risk categories, including content safety, CBRN (chemical, biological, radiological, nuclear), cybersecurity, persuasion and influence operations, child safety, self-harm, over-companionship, and regulatory compliance. Red teaming may span text, image, voice, and agentic model capabilities depending on project needs.
- This role requires creativity, curiosity, and an ability to think like an adversary while operating with strong ethical judgment.
- Craft creative prompts and multi-turn scenarios to stress-test AI guardrails across diverse risk categories
- Discover ways around safety filters, restrictions, and defenses using jailbreak, evasion, and prompt injection techniques
- Explore edge cases to provoke disallowed, harmful, or incorrect outputs
- Evaluate and score model responses against structured harm taxonomies and severity rubrics
- Document experiments clearly, including what you tried, why you tried it, and what it revealed
- Review and refine adversarial prompts generated by other team members
- Contribute to harm taxonomy development, calibration exercises, and inter-rater reliability work
- Collaborate with engineers, data scientists, and researchers to share findings and strengthen defenses
- Work with potentially disturbing content on a regular basis (see Content Warning below)
- Stay current on jailbreaks, attack methods, and evolving model behaviors
Будет плюсом
- Familiarity with Python or other scripting languages
- Experience working with LLM APIs or evaluation tooling
- Comfort with structured data annotation and rubric-based scoring
- Prior work in trust and safety, content moderation, QA, or security research
- Subject matter expertise in any high-risk domain (cybersecurity, chemistry, biology, medicine, law, finance, etc.)
Другое
- Location: Seattle, WA (candidates must reside in the Seattle metro area or be willing to relocate prior to start) Type: Contract, 40 hours per week
- Strong hands-on experience using multiple LLMs (ChatGPT, Claude, Gemini, open-source models, etc.)
- Intuition for crafting adversarial prompts; familiarity with jailbreak or evasion techniques is a strong plus
- Creative, adversarial problem-solving skills
- Clear and thoughtful written communication
- Strong ethical judgment and the ability to separate adversarial thinking from personal values
- Self-directed, collaborative, and comfortable in feedback-heavy environments
- Curiosity, persistence, and comfort with frequent failure in experimentation
- You treat every model response as a hypothesis to challenge
- You can switch between creative free-association and rigorous documentation in the same session
- You go deep into unusual interests (fandoms, niche internet cultures, gaming exploits, Wikipedia rabbit holes, etc.)
- You come from a creative background: writing, visual art, improv, puzzle design, or similar
- You are energized by finding the thing nobody else thought to try
- You are genuinely passionate about AI and follow the space closely
- This role involves regular and deliberate exposure to harmful content. You will encounter and intentionally generate content involving violence, self-harm, hate speech, sexually explicit material, child safety scenarios, and other categories of harmful output as part of structured adversarial testing. Candidates must be able to engage with this material professionally and sustainably. Support resources are available.