Researcher, Agent Safety, Training and Evaluations

OpenAI•San Francisco, CA

11d•$380,000 - $500,000•Hybrid

About The Position

The Agent Safety team works to ensure that increasingly capable AI agents act safely, exercise sound judgment, and remain aligned with user intent. Our mission is to reduce the probability of severe unintended outcomes from increasingly capable AI agents while preserving their ability to act effectively and autonomously. 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 autonomy (for example future versions of https://alignment.openai.com/auto-review/). We’re looking for strong executors with excellent judgment, comfort with ambiguity, and an understanding of frontier model research. You don’t need prior safety or alignment experience, we also welcome people that recently realized that alignment and safety is a critical area to contribute to. 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.

Requirements

Demonstrated strength in research engineering, ML engineering, quantitative research, or applied model research, with the ability to own ambiguous projects end to end.
Excellent technical execution across experimentation, data, evaluation, and/or infrastructure, plus strong intuition for modern frontier-model research.
Motivated by agent safety and eager to work on urgent, practical problems even if your prior work was outside safety or alignment.

Responsibilities

Train and evaluate frontier models to reduce harmful or misaligned agent actions, forming clear hypotheses and executing independently through ambiguity.
Mine incidents and build scalable measurement, data-processing, and evaluation systems that turn real failures into repeatable safety signals.
Collaborate closely with post-training, capabilities, oversight, and pre-training partners to ship research-backed mitigations into large-scale training and agent systems.