Adaptive Security — 80% of employees use unapproved AI tools, 12% of companies govern them

AI relevance: Unapproved AI tools connecting to corporate Workspace and Microsoft 365 via OAuth scopes create data-exfiltration paths that bypass network-layer controls and can feed sensitive context into unauthorized models.

  • Adaptive Security finds 80% of employees use unapproved generative AI tools at work, while only 12% of organizations have a formal AI governance policy — an 8-to-1 gap.
  • Most employees run 3–5 AI tools daily, with significant portions connecting to corporate data through OAuth tokens or browser sessions that never traverse the corporate network.
  • Three discovery surfaces dominate: OAuth connections to Workspace/M365, browser extensions with embedded AI features, and AI capabilities bundled into already-approved suites (Copilot, Gemini, Salesforce Einstein).
  • Browser-based AI tools authenticating via OAuth pull shared-drive contents through channels invisible to firewall logs and DLP-monitored email.
  • The critical finding: OAuth scopes persist after tools fall out of favor, and few organizations have routines for revoking stale third-party app authorizations.
  • The gap between approved-tool inventory and OAuth-app inventory represents the true shadow AI surface — the first number is what security teams think exists; the second is reality.

Why it matters

AI agents and coding tools increasingly connect to enterprise APIs through OAuth. Without governance, employees are granting third-party AI services read/write access to the same data pipelines that power authorized agent infrastructure — creating a parallel, unmonitored attack surface for data exfiltration and prompt-injection vectors.

What to do

  • Run a quarterly third-party OAuth audit against Workspace, Microsoft 365, and GitHub — pull the connected apps list sorted by permission scope and revoke unused authorizations.
  • Map approved AI tool inventory against actual OAuth app inventory; the delta is your program scope.
  • Implement OAuth consent screens that flag AI data-processing terms before employees grant access.

Sources: