GenAI Vulnerability Researcher (Remote, 35h/w)

Obsidian; San Francisco (CA); Remote; USD 130,000 - 160,000; Full time

Job summary

Obsidian is seeking a researcher to join its GenAI evaluation team, focusing on probing frontier models, identifying failure modes, and converting findings into stronger benchmark tasks. This remote role (about 35 hours/week) operates under Cincinnatus LLC as the employer of record.

Responsibilities include designing tasks, documenting results with reproducible steps, and collaborating with researchers to enhance benchmark quality.

Qualifications

• MSc or PhD in a STEM field or equivalent research experience requiring data analysis and coding.
• 1+ years of experience in a research, research-engineering, security, or AI-evaluation role.
• Demonstrated ability to identify vulnerabilities or failure modes in LLMs or ML systems through red teaming, adversarial testing, security research, or rigorous model evaluation.
• Proficiency in Python and Git, with the ability to script probes and analyses.
• Strong familiarity with LLM capabilities, limitations, and evaluation techniques.

Responsibilities

• Probe frontier AI models to reveal how they behave on coding, ML, and analysis tasks and find where they quietly fail.
• Design challenging tasks that are fair to grade and hard for models to exploit.
• Document findings with clear evidence and reproducible steps for others to follow.
• Collaborate with task authors to close loopholes, improve grading, and strengthen benchmarks.
• Share insights with researchers to continuously improve benchmark quality.

Skills

Red teaming

Adversarial testing

Model evaluation

Data analysis

Experiment scripting

Education

MSc or PhD in STEM

Tools

Python

Git

Job description

Obsidian is seeking a researcher to join its GenAI evaluation team, focusing on probing frontier models, identifying failure modes, and converting findings into stronger benchmark tasks. This remote role (about 35 hours/week) operates under Cincinnatus LLC as the employer of record.

Responsibilities include designing tasks, documenting results with reproducible steps, and collaborating with researchers to enhance benchmark quality.