About this role
Job title: Biological Safety Research Scientist
About the Role We are seeking biological scientists to help build safety and oversight mechanisms for our AI systems. As a Safeguards Biological Safety Research Scientist, you will design and develop safety systems to detect harmful behaviors and prevent misuse by sophisticated threat actors, shaping what responsible AI safety looks like in the biological domain.
What You'll Do
- Design and execute capability evaluations ("evals") to assess the capabilities of new models
- Collaborate with internal and external threat modeling experts to develop training data for safety systems, and with ML engineers to train these safety systems, optimizing for robustness and low false-positive rates for legitimate researchers
- Analyze safety system performance in traffic, identify gaps and propose improvements
- Develop rigorous stress-testing of safeguards against evolving threats and product surfaces
- Partner with Research, Product, and Policy teams to embed biological safety throughout the model development lifecycle
- Contribute to external communications, including model cards, blog posts, and policy documents related to biological safety
- Monitor emerging technologies for potential risks and mitigation strategies, and strategically address these
What We're Looking For
- A PhD in molecular biology, virology, microbiology, biochemistry, systems or computational biology, or related life sciences field, or equivalent professional experience
- Extensive experience in scientific computing and data analysis, with proficiency in programming (Python preferred)
- Deep expertise in modern biology, including both "reading" (high-throughput measurement, functional assays) and "writing" (gene synthesis, genome editing, strain construction, protein engineering) techniques
- Familiarity with dual-use research concerns, select agent regulations, and biosecurity frameworks (e.g., Biological Weapons Convention, Australia Group guidelines)
- Strong analytical and writing skills, with the ability to navigate ambiguity