How AI agents and models hold up on intelligence work, and where they break.
Preliminary results from a benchmark that tests AI agents on threat-infrastructure attribution. Across 400 runs, the agents never invented evidence, but some treated shared hosting as proof of common ownership.
