The hardest AI security problems now live in what an agent is permitted to do
Covered by 2 sources
Read the full article
OWASP's 2026 Top 10 for large language model (LLM) applications highlights excessive agency as a rising security risk, moving it from sixth to third place. The hardest AI security challenges now stem from what AI agents are authorized to do, rather than the model outputs themselves. Europe's Cyber Resilience Act mandates rapid vulnerability reporting for products but does not cover AI agent deployment, underscoring regulatory gaps.
