Research Engineer, Gemini Cyber Post-Training, DeepMind

Google•Mountain View, CA

1d•$147,000 - $210,000

About The Position

At Google, research-focused Software Engineers are embedded throughout the company, setting up large-scale tests and deploying promising ideas. They create experiments and prototypes, design new architectures, and work on real-world problems while staying connected to the research community through university partnerships and publications. As a Research Engineer in the Cyber Gemini team, you will advance research capabilities and build scalable infrastructure. The work covers cybersecurity models and custom orchestration harnesses, including skills and agent loops for vulnerability verification. You will develop datasets and post-training recipes to improve Gemini’s vulnerability-discovery capabilities, along with evaluation datasets to measure model performance. You will also work with DeepMind’s Cyber Strike initiative, a cross-organizational effort to build cybersecurity models for securing software and enterprise systems at scale. This includes contributing to finding, fixing, and exploit tracks; curating datasets; building agentic evaluation benchmarks; and integrating CodeMender’s vulnerability-discovery and patching capabilities into the Cyber Strike model. DeepMind is an interdisciplinary AI lab focused on advancing AI to address global challenges and support product innovation. The company describes public benefit, scientific discovery, safety, and ethics as priorities.

Requirements

PhD in Computer Science or a related field, or equivalent practical experience.
At least 2 years of industry or academic experience working in Reinforcement Learning environments.
Experience building distributed system infrastructure.
One or more publications in vulnerability research, such as auditing, reverse engineering, or exploitation.

Nice To Haves

Experience with program analysis techniques, such as static analysis, symbolic execution, or fuzzing.
Background in machine learning and statistics.
Independent and proactive attitude, and a passion for AI safety and security.

Responsibilities

Build datasets, including RL environments and supervised fine-tuning (SFT) data, and iterate on post-training recipes.
Debug and analyze RL training runs to identify bottlenecks and improve training efficiency and effectiveness.
Build evaluation pipelines, benchmarks, and leaderboards to measure agent performance.
Build scalable infrastructure and custom orchestration harnesses, including skills and agent loops for vulnerability verification.
Develop datasets and post-training recipes to improve Gemini’s vulnerability-discovery capabilities, and create evaluation datasets to measure model capabilities.
Contribute to the finding, fixing, and exploit tracks of the Cyber Strike initiative; curate datasets, build agentic evaluation benchmarks, and integrate CodeMender’s vulnerability-discovery and patching capabilities into the Cyber Strike model.