Dylan Sam

This stream will focus on safety and alignment pretraining, data poisoning, and the role of harmful data in training safer models.

Apply

View all streams

Stream overview

I'm interested in mentoring projects in several directions:

Safety or alignment pretraining: There is a large variety of topics in this direction, including ideas in generating synthetic pretraining data to help with improving robustness or alignment with human values, compliance with safety policies.
Data poisoning: As synthetic data becomes an increasingly large portion of pretraining corpora, new and more subtle forms of data poisoning become possible. I'm interested in projects that develop methods for detecting, characterizing, and defending novel forms of such attacks.
The role of harmful data in building safer models: There is growing evidence that retaining some harmful content during pretraining, rather than filtering it all out, can improve a model's ability to reason about harms and ultimately become safer after post-training. I'd like to mentor work that deepens our understanding of this phenomenon: when does exposure to harmful data help vs. hurt, and how can we design pretraining pipelines that leverage this insight responsibly?

Mentors

Dylan Sam

OpenAI

Member of Technical Staff

SF Bay Area

—

AI Control and Monitoring

Alignment Training Methods

Dylan is a safety researcher at OpenAI, where he works on curating better/safer training data and monitoring models for harmful behavior.

Before that, he completed a PhD in the Machine Learning Department at CMU.

Fellows we are looking for

Project selection

The Winter 2026 cohort offers a wide range of research streams led by experts across AI alignment, interpretability, governance, and safety. Each stream provides its own research agenda, methodology, and mentorship focus.