About the job
The Alignment and Reliability division investigates and creates auditing frameworks, defensive safeguards, specialized toolsets, and autonomous agents to upgrade GDM’s flagship foundation systems. The mandate for the Research Scientist/Engineer centers on engineering novel heuristic and statistical mechanisms to elevate user-facing architectures. In this role, you will be a technical contributor capable of advancing fresh investigative hypotheses from conception to live release with full lifecycle accountability, balancing experimentation with rapid incident mitigation. Our group specializes in enhancing the ethical integrity and alignment standards of machine intelligence.
Responsibilities
Drive innovation and understanding of safety and security jailbreaks, owning the problem and delivering mitigations which can be deployed at scale in partnership with product areas.
Develop red and blue teaming methods for frontier GenAI models spanning text-to-text, multimodal, and agentic capabilities, delivering actionable insights and solutions.
Explore data, reasoning and algorithmic solutions to make sure Gemini Models are safe, maximally helpful, and work for everyone.
Improve Gemini’s adversarial with a focus on high-stakes abuse risks in agentic settings.
Develop and execute experimental plans to address known gaps, or construct entirely new capabilities.
Qualifications
Minimum
PhD degree in Computer Science, a related field, or equivalent practical experience.
2 years of experience in Large Language Model safety or security.
Preferred
Experience in developing and leveraging agentic workflows around safety, behavior, and alignment.
Experience with synthetic data generation pipelines, building evaluations and mitigations for non-verifiable tasks using methods such as LLM-as-a-judge, rubric-based rewards, etc.
Experience taking research from concept to product.
Experience with collaborating or leading an applied research project.
Strong experimental taste with good judgment regarding baselines, ablations, and what is worth testing.
Track record of publications at NeurIPS, ICLR, ICML.