Researcher, Multimodal Safety

About the Role

As a Researcher on the Chat and Multimodal Safety team, you will help shape how frontier models perceive and reason the world, and translate that understanding into safe behavior.

We’re looking for people who combine deep technical expertise with strong safety judgment. Strong candidates often bridge perception and language: they may have built vision-language models, worked on modality fusion or image encoders, developed multimodal post-training or evaluations, or advanced safety for image, video, or audio systems.

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

• Define and advance multimodal safety research for text, vision, and audio, connecting perception and semantic understanding to safe model behavior.
• Build training and evaluation methods for VLMs, including post-training, safety evals, and interventions that help models respond safely and appropriately in varied contexts.
• Collaborate closely with Personal AGI, Consumer Devices, and product/model teams to translate research into safer ambient, embedded, and personalized multimodal experiences.

You might thrive in this role if you

• Have a track record of building or advancing multimodal models, with depth in vision-language models, video understanding, image generation, audio, or multimodal reasoning-and fluency across both perception and language.
• Understand how multimodal systems work end to end, from encoders, projection layers, and modality fusion to cross-modal reasoning, scaling, and inference tradeoffs.
• Have improved frontier model behavior through post-training, using approaches such as SFT, RL, data curation, synthetic data, evaluation, and rigorous error analysis.
• Bring strong research and engineering judgment to open-ended safety problems: you can form testable hypotheses, design decisive experiments and evaluations, diagnose model failures, and translate findings into robust improvements.