Zero Trust for Agentic AI: Never Trust an Agent, Always Verify Its Actions
Security Boulevard, Wednesday, October 28th, 2026
Applying zero trust to agentic AI means verifying every agent action through policy checks, scoped permissions and an execution gateway.
Swapneswar Sundar Ray argues that agentic AI breaks the assumption that people specify every action in advance, because agents reason their way to actions such as issuing refunds, and since reasoning is not authorization, proposed actions should pass a separate policy check before executing.
He recommends task-scoped, just-in-time permissions instead of standing broad access and treating every tool call as a security checkpoint.
Because prompt injection cannot be fully solved, systems must assume manipulation and prevent it from triggering real actions.
The suggested safeguard is an execution gateway that checks identity, evaluates policy, scores risk, and requires human approval above thresholds, while keeping systems auditable and recoverable.