The New Blind Spot in Enterprise Security

Corporate security teams have long fought "Shadow IT" — employees quietly adopting unsanctioned laptops or cloud services to work around bottlenecks. Firewalls and policies were built to hunt down those rogue tools. But a different, far more subtle problem has emerged.

Consider a developer on a tight deadline who pastes proprietary code into a public AI chatbot for refactoring suggestions. The revised code comes back quickly; the engineer moves on. What they don't realize is that the query — including the code snippet — may now sit in the AI provider's logs or training data. A piece of the company's intellectual property has left the organization, silently and without oversight.

That scenario is playing out across enterprises daily. Employees are feeding confidential documents into LLMs for summaries, generating marketing copy, and writing code. Traditional security tools weren't designed to interpret what happens when a browser tab talks to a large language model. This unmanaged usage — "Shadow AI" — is a high-stakes visibility gap that demands a new kind of control: one that doesn't cripple the productivity gains these tools offer.

Visibility Before Enforcement

Cloudflare's answer starts not with blocking but with understanding. The Shadow IT Report within Cloudflare One is an analytics tool that gives security teams a data-driven view of application usage across their organization, derived from actual traffic insights rather than guesswork.

The report categorizes application activity and can be filtered specifically for AI services. The classification — powered by Cloudflare's network and threat intelligence — covers general-purpose chatbots like ChatGPT, code assistants like GitHub Copilot, and specialized creative or analytical tools like Leonardo.ai. Teams can see not just that an AI app is in use, but precisely which one and by whom.

Rebuilding the Analytics Engine

Cloudflare Gateway has logged some application usage data for some time, but the sheer scale — hundreds of millions of rows per day for the largest customers — created slow queries on longer time windows. The original implementation also lacked the customization and filtering required to properly investigate AI-specific usage.

The fix required a cross-team overhaul of the analytics platform. The team adopted TimescaleDB, as detailed in a July 2025 engineering post, to aggregate and compress long-term data, dramatically improving query performance. Traffic data from Gateway gets stored in a Timescale database. From there, specific materialized views were built around the Shadow IT and AI use case, centered on the answers security teams actually need:

  • Which users are accessing unapproved applications?
  • How much bandwidth are those apps consuming?
  • Are end-users in unexpected geographic regions interacting with unreviewed tools?
  • Which devices are the heaviest bandwidth consumers?

The team applied a consistent analytics framework — filterable timeseries graphs and top-n views with drill-down capabilities — to surface the application-level detail customers had been asking for.

A Four-Step Workflow for Taming Shadow AI

1. Proxy Traffic Through Gateway

Cloudflare Gateway acts as the inline proxy for all organizational traffic, wherever users are located. When an employee reaches an AI application, the traffic passes through Cloudflare's global network, where hostnames are inspected and mapped to application definitions. TLS inspection is optional for Gateway customers but required for the analytics to function.

Every interaction is logged with user identity, device posture, bandwidth consumed, and geographic location — the context needed to understand who is using which AI tools, when, and where.

2. Review Dashboards and Drill Down

All that data appears in the Shadow IT Report inside the Cloudflare One dashboard. Filtering for AI applications reveals two levels of insight:

  • High-level overview: See the leading AI applications in use, overall adoption trends, and data volumes — enough to focus governance efforts.
  • Granular drill-downs: Click into any application to view the specific users or groups accessing it, usage frequency, locations, and bytes transferred.

3. Categorize with Approval Statuses

Not all AI tools carry equal risk, and the framework acknowledges that with a formal three-tier approval system:

  • Approved: Passed internal security vetting and officially sanctioned.
  • Unapproved: Concerning privacy policies, known vulnerabilities, or misaligned with business goals.
  • In Review: Gray-area or newly discovered tools flagged for acknowledgment while due diligence proceeds.

4. Enforce Automatically

Approval statuses feed directly into Gateway policies, enabling automatic enforcement on Cloudflare's edge network:

  • Block unapproved AI with a simple HTTP policy that halts traffic to flagged applications.
  • Apply DLP to "In Review" apps so that sensitive data — credit card numbers, PII, internal project names, source code — is blocked from being included in prompts, with a notification explaining why.
  • Block file uploads only to allow basic interactions while preventing mass data egress.
  • Route sessions through Cloudflare Browser Isolation, which executes the session remotely, giving you control over file uploads, clipboard actions, and keyboard input while an application is being evaluated.
  • Audit approved applications by logging interactions for compliance and applying data-handling rules.

That structured workflow makes it possible to audit AI usage regularly and update policies as risk profiles evolve.

Forensics in Log Explorer

Dashboards surface the big picture, but incident investigations demand deeper digging. Cloudflare Log Explorer provides raw, unsampled logs with full context, queryable directly in the dashboard or via API — eliminating the need to push massive log volumes into third-party SIEMs for every query.

Pre-populated SQL queries support Shadow AI investigations, including:

  • Tracing user activity when the report flags heavy usage on an unreviewed application — query by user, application category, or specific AI service.
  • Analyzing exfiltration attempts by correlating DLP matches with AI application categories to see precisely what data was being transmitted.
  • Spotting anomalous usage patterns — spikes from a single IP address or unexpected geographic origins that could signal compromised credentials or policy evasion.

For Cloudflare One customers, the Shadow AI Report is available now. Security teams can find it in their dashboard to begin mapping AI adoption across the workplace.