Open science and tools for white-box auditing of frontier AI agents.
Black-box and white-box methods offer complementary views of agents' behaviour. Both are essential for frontier AI audits.
MORE ABOUT US →Our research and open infrastructure are co-developed with organisations at the forefront of agentic evaluations.
We study how agents encode latent beliefs, goals, plans, and their impact on model behaviours at scale.
We develop cognitive map probes recovering an agent's approximate beliefs about its environment, and use them to prototype white-box evaluations explaining suboptimal actions in light of the agent's imperfect beliefs in simple 2D gridworlds.
A technical report on failure cases of stateful agentic systems based on the OpenClaw harness when interacting in a complex environment including productivity and messaging tools, other agents and multiple human red-teamers.
In this critical review about recent claims about misaligned goals in LM agents, we argue in favor of a research programme grounded in robust theoretical frameworks and better reporting standards.
An experienced team working on interpretability and frontier AI evaluations.
