How Architectural Controls Help Can Fill the AI Security Gap

This Dark Reading interview with NCC Group's AI/ML security lead highlights research showing that current AI guardrails are insufficient: poorly constrained LLMs can be coerced into executing arbitrary code, exfiltrating passwords, and exposing databases. The discussion recommends shifting from object-based to data-based permissions, applying architectural controls (never expose high-privilege AI to untrusted data), using capability shifting, and integrating LLM data flows into threat modeling to sever potential attack chains.