Our research program is organized around three pillars. Each one starts from the same place: an explanation that no longer holds.
How do models and agents actually behave, compared with how they are documented to behave?
We measure behavior across models, versions, and settings, and track how it drifts. That includes compliance, refusal, deception, sycophancy, and the gap between stated reasoning and action.
Where do guardrails, defenses, and instructions break under deliberate pressure?
We attack defenses methodically, as an attacker would, and record every attempt. The result is a map of failure classes, not a collection of clever prompts.
What happens when autonomous systems act in real environments, including their own?
Agents with tools, memory, and permissions behave differently from chat models. We study what they do when given room: how they route around restrictions, and how they report on their own work.