About the Role
As a Research Scientist working on Agent Robustness, you will work on the fundamental challenges of building AI agents that are safe and aligned with humans. Scale Labs has launched a new team focused on policy research, bridging the gap between AI research and global policymakers to make informed, scientific decisions about AI risks and capabilities. This team collaborates broadly across industry, the public sector, and academia and regularly publishes its findings.
Responsibilities
- Research the science of AI agent capabilities with a focus on how they relate to safety, risk factors, and methodologies for benchmarking them.
- Design and build harnesses to test AI agents’ tendency to take harmful actions when pressured to do so by users or tricked into doing so by elements of their environment.
- Design and build exploits and mitigations for new and unique failure modes that arise as AI agents gain affordances like coding, web browsing, and computer use.
- Characterize and design mitigations for potential failure modes or broader risks of systems involving multiple interacting AI agents.
Requirements
- Commitment to our mission: Promoting safe, secure, and trustworthy AI deployments as frontier AI capabilities continue to advance.
- Practical research experience: Comfortable building and leveraging agent scaffolding, designing evaluation harnesses, and quickly turning new ideas from the research literature into working prototypes.
- Experience with post-training and RL techniques: Such as RLHF, DPO, GRPO, and similar approaches.
- Track record of published research: In machine learning, particularly in generative AI.
- Experience: At least three years addressing sophisticated ML problems, whether in a research setting or in product development.
- Communication skills: Strong written and verbal communication skills to operate in a cross-functional team.
Nice to Have
- Hands-on experience with agent evaluation frameworks such as SWE-bench, WebArena, OSWorld, Inspect, or similar tools.
- Experience with red-teaming, prompt injection, or adversarial testing of AI systems.