"Across most organizations today, employees are running three to five AI tools on any given day." That simple fact, lifted from the reporting, is the clearest signal of a widening operational problem: useful tools are multiplying faster than the controls meant to govern them.
Three technical vectors that create the shadow AI gap: OAuth, browser extensions, and bundled AI
The piece identifies three areas that account for the majority of shadow AI activity. First, OAuth connections: "Most AI tools request access to Google Workspace or Microsoft 365 through OAuth," granting read or write permissions to corporate data and exposing shared drives, emails, and internal documents. Second, browser extensions: many tools run as extensions and "never touch the operating system," so traditional endpoint management tools miss them. Third, AI features bundled inside already-approved tools — examples named are Microsoft Copilot, Google Gemini, and Salesforce Einstein — which may have been introduced after an original vendor review and without a separate security evaluation. Together, these vectors let AI activity bypass network-era controls.
Step 1: Build an inventory — OAuth audits, browser scans, and employee surveys
A security program "can only manage what it can see." The recommended start is a comprehensive inventory of every AI tool in use, who is using it, and what data it can access. Practical techniques listed include a quarterly audit of connected third-party apps sorted by permission scope (to find OAuth grants), a browser management solution or a lightweight agent to scan for active extensions, and an employee survey. The report notes that "many shadow tools surface through surveys that automated discovery misses entirely."
Step 2 and 3: Policies that work with employees, and a fast lane for new tools
Policies that simply list prohibited tools tend to fail; the guidance here is to produce a practical AI governance policy that identifies approved tools and outlines a clear, rapid process for requesting new ones. An effective policy should include five elements: a current list of approved tools and where to find them; clear data classification rules (including categories such as customer records, source code, and financial information that "should never be entered into any AI tool"); a verified data training opt-out status for each approved tool; a defined process for requesting new tools with a target turnaround time; and a plain-language explanation of why the guidelines exist.
To prevent shadow adoption, the report urges a "fast lane" for lower-risk requests: most AI tool requests do not warrant a full procurement review, and a structured intake form with defined evaluation criteria can shorten turnaround times. Suggested evaluation criteria include data access scope, vendor security practices, data training opt-out status, compliance certifications, and whether a functional equivalent already exists on the approved list.
Step 4 and 5: Browser-native monitoring and just-in-time coaching as a combined safety layer
Continuous, browser-native monitoring gives security teams "the real-time picture they need" without rerouting employee traffic, and it delivers a protection signal back to employees when a tool may be putting credentials or company data at risk. The monitoring signals should feed into an employee's broader risk profile — alongside phishing simulation results and training completion data — because "risky behaviors compound."
Making secure choices the easiest path requires two behavioral tools: just-in-time coaching and reasoning-focused training. Just-in-time coaching delivers a brief, contextual prompt at the moment an employee attempts to use an unsanctioned tool — "a well-designed prompt" should tell the employee the concern, direct them to an approved alternative, and take less than thirty seconds to read. Training should explain the reasoning behind the rules so employees can apply judgment to future tools and threats; for example, an employee who understands that "OAuth connections to corporate Google Workspace can expose the entire shared drive to a third-party vendor" will apply that logic to new tools they encounter later.
What this means for security teams, procurement leaders, and employees
- Security teams: build discovery capabilities (OAuth audits, browser scans, and surveys), publish an approved-tools list, and adopt browser-native monitoring so they can see activity that bypasses network controls.
- Procurement and IT reviewers: implement a structured intake form and documented evaluation criteria to enable a faster, lower-friction approval path for low-risk tools while reserving detailed reviews for higher-risk cases.
- Employees: expect clearer guidance and faster approvals if organizations publish approved tools and provide just-in-time coaching; conversely, slow approval processes will continue to drive workarounds.
The report’s central thesis is straightforward: AI adoption is an operational benefit, not a policy failure, but it requires practical channels to remain safe and visible. Organizations that combine full discovery, sensible policy, a fast approval lane, browser-native monitoring, and just-in-time coaching "tend to handle it best." The piece closes by pointing to a vendor solution — Adaptive Security's AI Governance product — which it says provides real-time visibility, automated policies, and just-in-time coaching; more information is available at adaptivesecurity.com.
Read the original story: https://thehackernews.com/2026/05/5-steps-to-managing-shadow-ai-tools.html
