Hristo Todorov

I’m currently interning at Armadin, building robust evals to identify emerging cyber threats.

I just graduated from Stanford, where I studied math and ancient history.

Research

I want to uncover the general principles behind how large-scale AI systems work, and help ensure they are secure and robust. I’m interested in problems like cyber capabilities, backdoors, robust sandboxing, and mechanistic interpretability.

I worked on Archon (ICML 2025), a framework for test-time compute, in Chris Ré’s HazyResearch, and did research on robotics in Dorsa Sadigh’s ILIAD lab.

History

I often think about the Roman empire. Recently, I have been trying to figure out:

• Why did Rome never industrialize? Why were the Romans so bad at innovating? Why didn’t their patronage networks cultivate more “scientists” in the provinces?
• What were the non-obvious institutions, formal or informal, that let people across the empire collaborate?

I wrote an essay on how collegia, Rome’s voluntary associations, quietly “Romanized” the empire, and provided the structural scaffolding for the early Christian church.

Personal

I grew up in the cherry capital of Bulgaria. Studying math there built a deep appreciation for simple, intuition-heavy “tricks” for solving problems, a philosophy that still shapes the open questions in AI and science I care about. It’s also where I did some of my first research, on adversarial robustness, back in high school.

In my free time, I love reading books, skiing, and playing tennis.