Researcher, Agent Safety, Oversight and System Mitigations

Our work spans three areas

• Training: Create training methods, environments and data that teach agents to make better decisions in consequential situations. We turn real-world failures into training signals that prevent similar incidents, and identify precursor behaviors and mitigations to address emerging risks.
• Measurements: Build evaluations and production metrics that identify emerging risks and measure whether our interventions work.
• Oversight: Develop oversight and system mitigation mechanisms that reduce harmful actions while preserving useful agent autonomy (for example future versions of auto-review).

About the Role

This role focuses on oversight and system-level mitigations that enable increasingly capable agents to operate safely and autonomously in real environments. We also study longer-term questions about how increasingly capable agentis systems can be supervised, constrained, and corrected.

We’re looking for a safety&security minded researcher or engineer who can reason rigorously about security boundaries and agent behavior, then build and test practical mitigations. A background in AI control or security is welcome but not required.

This role is based in San Francisco, CA. We use a hybrid work model of 3 days in the office per week and offer relocation assistance to new employees.

In this role, you will

• Design, build, and evaluate system-level controls for agent actions like agent-based review. Plan how they fit in a broader system including sandboxing with process isolation and permission boundaries.
• Work closely with a Codex harness engineering team to productionize the AI controls.
• Red-team end-to-end agentic systems to measure whether controls prevent data exfiltration, unsafe tool use, and other harmful outcomes.
• Improve the safety-productivity tradeoff by measuring and reducing missed harmful actions, unnecessary blocks, approval burden, and latency.

You might thrive in this role if you

• Have strong systems or security instincts and can reason concretely about isolation boundaries, permissions, attack surfaces, and failure modes in complex systems.
• Enjoy turning ambiguous safety questions into concrete threat models, reproducible experiments, and practical mitigations, and revising your approach based on evidence from deployment.
• Can build robust experimental infrastructure and design evaluations that distinguish promising mitigations from brittle ones.
• Are deeply interested in frontier AI alignment, safety and control.