- 经验
- 5年以上
- 薪水
- USD 70 – USD 84 / hour
- 职位空缺
- 1
- 发布
- 4小时前
- 工作模式
- 在办公室
- 学历
- 学士学位
- 恢复
- 需要申请
你的工作地点
职位描述
About the Role
We are seeking experienced professionals to join as AI Safety Red Teamers. This position involves crafting challenging adversarial prompts to rigorously test advanced AI models, uncovering vulnerabilities, and assessing AI behaviors in complex and high-risk scenarios often involving ambiguous or "grey-area" topics.
Key Responsibilities
- Create adversarial prompts designed to push frontier AI models to their limits and identify potential failure points.
- Detect unsafe behaviors including jailbreak attempts, hallucinations, policy violations, and other weaknesses within AI systems.
- Assess model resilience in sensitive domains such as misinformation, cybersecurity, biosecurity, fraud, and political topics.
- Thoroughly document identified vulnerabilities and contribute to the compilation of safety benchmarks and red-teaming reports.
- Work collaboratively with AI researchers to enhance model alignment, robustness, and overall safety.
Candidate Profile
- Must hold a bachelor's degree or higher in fields such as Computer Science, Cybersecurity, Journalism, Communications, Psychology, Biology, Chemistry, Public Policy, or a related area.
- At least five years of professional experience in any relevant sector including AI Safety, AI Red Teaming, Trust & Safety, cybersecurity, investigative journalism, life sciences, or related disciplines.
- Possess strong analytical and reasoning capabilities, exceptional prompt design expertise, and excellent written communication skills.
- Proven experience in designing adversarial prompts or evaluating cutting-edge AI systems is essential.
Preferred Qualifications
- Background or experience in AI Red Teaming, Reinforcement Learning with Human Feedback (RLHF), Supervised Fine-Tuning (SFT), AI Alignment, or Trust & Safety domains.
- Knowledge of jailbreak testing, prompt engineering techniques, or adversarial evaluation practices.
- Specialized understanding of one or more sensitive domains such as cybersecurity, biosecurity, political content, misinformation, or scientific safety.