Caleb DeLeeuw

Caleb DeLeeuw is an independent AI safety researcher and Executive Director of Copyleft Cultivars, an open-source bio research nonprofit. His first-authored AAAI 2026 paper, The Secret Agenda, found that auto-labeled sparse autoencoder features for deception rarely fire while a model is lying, and in later work refined this method. He has participated in multiple Apart Research hackathons. He's trained and published over 350 SAEs, as well as the first natural language autoencoders released outside Anthropic and wrote NLAttack, the first open benchmark for whether NLAs faithfully report a model's internal state in realistic contexts, which his current research is extending across more model families.

Projects

No grants recorded.