USC paper shows AI agents leak secrets via the "mosaic effect" without reading classified files
2C_ornot2C · x · 2026-11-09
A new USC paper (Agent Mosaic, Oct 21) demonstrates the "mosaic effect": each document an agent can access looks harmless, but the agent can assemble a sensitive conclusion from pieces and say it aloud. Across 30 synthetic cybersecurity, medical, and military scenarios, the best-performing model disclosed forbidden composites in 7.08% of simulated answers even on benign single-document questions — worse in live harnesses. Adding classification headers and aggregation rules still left 32.37% unsafe disclosures, and refusals often reveal the prohibited conclusion anyway. Per-document permissions aren't enough; the real question is what an agent can infer from everything it touches.
More from Safety
- AI cameras on trash trucks are reporting homes to the city, triggering hundreds of code warnings — JarnoDuursma · 2026-11-09
- Stanford's Evo AI designs 16 working bacteria-killing viruses, sparking access debate — coursiv_ · 2026-11-09
- Veteran pentester: AI agent swarms get blocked fast — breaches still come from bad passwords — ziv_ravid · 2026-11-09
- Hermes Agent ships Security plugin: parallel subagents find and verify code vulnerabilities — NousResearch · 2026-11-09
- Upcoming post on AI safety in China to be published tomorrow — JacquesThibs · 2026-11-09
- Banning AI Is Like Banning Uranium Enrichment, Safety Researchers Quip — DavidSKrueger · 2026-11-09