AI Agent Security Must Start With Visibility, Not Enforcement

Summary

Organizations should inventory and continuously monitor AI agents before applying Zero Trust controls. The article outlines visibility gaps across shadow AI, encrypted network traffic, and rapidly changing agent deployments, and recommends distinct agent identities and

Key points

• The article argues that agent discovery and inventory must come before enforcement controls, since controls cannot govern agents an organization cannot see.
• It cites Veeam research reporting that 70% of organizations have AI workflows touching sensitive corporate data without full oversight, while 67% say IT cannot fully track employee-built autonomous workflows.
• At METR, an attacker reportedly found an employee’s personal EC2 instance running an agentic app, bypassed authentication, and obtained its model-provider API key; token use reached the equivalent of $600,000 over three weeks.
• A single monitoring source is insufficient: encrypted network traffic, browser-based tools, endpoint telemetry, and SaaS-embedded agents each leave visibility gaps.
• The article recommends correlating network metadata, endpoint and browser telemetry, identity records, and SaaS logs to build an agent inventory.
• Continuous monitoring is needed because agents and short-lived clones can be deployed and removed faster than periodic audits can track.
• It recommends giving each agent a distinct identity, task-scoped permissions, and an audit trail that records tool calls and actions.