Enterprises systematically overestimate their visibility into AI agents (Cloud Security Alliance, Apr 2026: 82% had discovered at least one AI agent running without their security or IT team's knowledge in the past year while 68% believed they had strong visibility, with only 21% running any formal agent decommissioning process), and because a written policy cannot be enforced against agents nobody can see, continuous discovery rather than policy is the binding first control.

Anchored on the CSA + Token Security survey 'Autonomous but Not Controlled: AI Agent Incidents Now Common in Enterprises' (published 21 Apr 2026, n=418 IT/security professionals, fielded Jan 2026): 82% found at least one unknown AI agent in the past year, 68% believe they have strong visibility, 65% had an AI agent security incident, of those incidents 61% involved data exposure, 21% run a formal decommissioning process; Itamar Apelblat (CEO/Co-Founder, Token Security) quote. Scale anchored on CrowdStrike press release (23 Mar 2026): 1,800+ distinct AI applications across ~160 million unique instances, Shadow AI Discovery for Endpoint capability; Michael Sentonas (President) quote. VERIFIED 2026-06-08 via cloudsecurityalliance.org press release/artifact and crowdstrike.com press release. PRECISION: the 61% data-exposure figure is of the incident group (65%), not all 418 respondents — stated as such in the body. Distinct from AM-204 (NHI governance vacuum — machine-identity scale/lifecycle) and AM-168/shadow-ai-discovery-playbook (broad shadow AI): this piece's core is the believed-vs-actual visibility gap and discovery-as-first-control. 90-day cadence. Triggers: (1) a later large-sample survey showing find-rate and believed-visibility converging; (2) discovery tooling becoming a default in major endpoint/cloud platforms; (3) incident data showing policy maturity, not discovery, separates breached from unbreached. Siblings: AM-204 (NHI governance vacuum), the shadow-AI discovery playbook, approved-tool-unapproved-capability.