Why transparent AI agents matter more than you think

This report warns that prompt injection against AI agents is an escalating enterprise risk—highlighting a Snyk audit (36% of skills with critical issues) and a Mozilla proof-of-concept where indirect prompt injection caused Claude to spawn a reverse shell—and recommends governance, bounded tenant awareness, strict access controls, standardized telemetry, and detection/containment controls (UEBA, NDR, multi-layer AI filtering) to prevent catastrophic data loss and unauthorized system actions.