6 Ways Teams Can Leverage Zero-Trust to Manage AI Agents
SC Media, Tuesday, September 8th, 2026
Cato Networks' Etay Maor argues AI agents should get zero-trust treatment and offers six guidelines to limit their blast radius.
Etay Maor of Cato Networks writes that 62% of organizations are adopting agentic AI while 53% report agents exceeding intended permissions, citing agents that wiped a developer's drive and deleted a production database.
He traces failures to unverified trust: reward hacking, confusing data with instructions, literal execution and excessive permissions, with human approval weakened by fatigue. His six guidelines are to treat content as untrusted input, enforce least privilege outside the model, re-authenticate high-risk actions, design sensible human approvals, add behavioral monitoring, and diversify the model stack.