Only 11% of production agents pass the AI agent security bar

The AIRQ Q2 2026 assessment of 100 AI agents finds nearly universal exposure to a "lethal trifecta" (private data access, ingestion of untrusted content, and outbound actions), with coding and computer-use agents showing the largest blast radii and weakest defenses; the report recommends sandboxing, stronger procurement/configuration controls, independent verification of claimed defenses, and regular re-audits.