Empirical security research on emerging AI systems.

We study how AI systems actually behave under real conditions, and publish the methods alongside the findings.

Most accounts of AI capability and risk are anecdotal or vendor-supplied. EEP exists to replace that with measurement. We construct concrete threat models, run open and frontier systems against them, and document the results in full — including where the experiment failed or surprised us. The goal is not a verdict on whether a model is “safe,” but a precise account of which controls are actually load-bearing.

01

Empirical

Every claim is grounded in data we collected. We report the measurements, including negative results, rather than the narrative.

02

Reproducible

Harnesses, prompts, and code are published with each report so peers can re-run, challenge, and extend the work.

03

Policy-legible

Findings are mapped to established taxonomies and translated into language a decision-maker can act on.