AI agents, interpretability, and biosecurity Projects Divergent soft prompts converge on personas Contextualized soft prompts are interpretable Socials Email GitHub LinkedIn Google Scholar Twitter Resume