70% of Organizations Run Untracked AI Agents on Sensitive Data, METR Breach Shows $600k Token Theft via Uninventoried EC2

The core failure is inventory absence before any Zero Trust controls. Organizations deploy autonomous agents that reach model APIs and internal data stores yet lack named owners or scope definitions. The METR incident revealed an employee vibe-coded agent on a personal instance whose authentication bypass allowed token exfiltration; internal dashboards ignored rate-limited request volume. This matches patterns from prior cloud shadow IT where spend monitoring preceded policy enforcement.

SANS Zero Trust for AI Agents checklist explicitly sequences inventory ahead of authorization layers because an unenforced proxy has no population to govern. Veeam's parallel finding that 67% of IT teams cannot track employee-built workflows indicates the same adoption-first dynamic seen in 2015-2018 cloud migrations. Finance and procurement records of API key issuance and token spend function as discovery signals that predate technical telemetry.

Treating AI agents as critical data paths rather than consumer tools exposes the mismatch: consumer electronics supply chains already require SBOMs for safety-critical components, yet enterprise AI deployments operate without equivalent asset registers. The result is lateral movement risk from personal instances into production model providers.

Next steps require procurement-mandated key issuance logs and automated discovery of agent-linked cloud resources within 90 days, followed by owner assignment before any authorization policy is applied.