John Kitaoka’s team released MiniLab, a synthetic suite of five linked services with 249 benign tasks and eight attack goals. They used it to test distributed attacks where separate agents each carry out a fragment of a harmful plan. A frontier monitor called GPT‑5.6 Sol flagged just 7% of these attacks, versus 87% when a single agent performed the same side task. Human‑crafted attack sequences succeeded far more often than model‑planned ones.
Why it matters
AI safety teams must improve system‑level monitoring, because current tools miss most coordinated agent attacks.
