A mid-size SaaS company with 800 employees had a standard security stack: a web proxy, a CASB, endpoint detection and response, and a SIEM. They believed they had visibility into their AI tool usage. When they deployed LeakSnitch's browser extension across their organization, the dashboard revealed 14 different AI tools in active use. Only 2 of those tools were IT-approved. The other 12 had no DLP, no audit trail, and no data retention policies.
In the first month, the extension detected 347 PII exposures across those unapproved tools. Credit card numbers in Perplexity prompts. Database connection strings in a free AI code assistant. Patient data in a shadow AI note-taking application. None of these exposures would have been caught by the existing security stack because the data was traveling through HTTPS request bodies, not file uploads or email attachments.
This is not an outlier. It is the norm.
The Scale of Shadow AI
Industry surveys from Gartner and Netskope estimate that 70-80% of employees use AI tools that are not sanctioned by their IT department. The most common unapproved tools fall into four categories:
AI coding assistants. Developers install Copilot, Cursor, Codeium, Cody, and Continue directly into their IDEs. These tools send code context to cloud APIs. When a developer pastes a production config file or a database connection string into a chat prompt, that data leaves the corporate network through a channel that no web proxy or CASB can inspect at the content level.
GitHub Copilot
Cursor
Codeium
Cody
Continue
AI note-taking and productivity tools. Tools like Notion AI, Mem, and Otter.ai are installed by individual employees who want AI-powered summaries and search. These tools process meeting transcripts, internal documents, and project plans through cloud models. The data is stored on the tool provider's infrastructure, outside the enterprise's data governance boundary.
General-purpose AI chat platforms. ChatGPT, Claude, Gemini, and Perplexity are the most visible shadow AI tools. Employees use them for research, writing, analysis, and problem-solving. The data they paste into these platforms includes internal business logic, customer data, and strategic plans.
ChatGPT
Claude
Gemini
Perplexity
Specialized AI tools. Industry-specific AI tools for legal research, medical coding, financial analysis, and marketing content generation are adopted by individual teams without IT review. Each tool has its own data handling policy, most of which are not reviewed by the enterprise's legal or security teams.
Why Traditional DLP Misses Shadow AI
Traditional data loss prevention tools were designed for a world where data left the organization through email, file uploads, and USB drives. They inspect SMTP traffic, HTTP file uploads, and endpoint file system activity. They do not inspect the request body of an HTTPS POST to an API endpoint.
AI tools communicate through API calls. The user types a prompt, the browser sends a POST request with the prompt text in the request body, and the AI tool responds with generated text. To a web proxy, this looks like normal API traffic. There is no file attachment. There is no email recipient. There is no USB device. The data leaves the organization through a channel that was designed to look benign.
This is why the SaaS company's existing security stack missed 347 PII exposures in one month. The data was not being exfiltrated through traditional channels. It was being sent through API calls that looked identical to legitimate AI tool usage.
The Cost of Shadow AI
The direct costs of shadow AI are measurable: data breach notification costs, credential rotation overhead, incident response time, and compliance penalties. But the indirect costs are larger.
Loss of data governance. When data flows through unapproved AI tools, the enterprise loses the ability to control where that data is stored, how long it is retained, and who has access to it. This creates compliance risk under GDPR, HIPAA, PCI DSS, and other regulatory frameworks.
Erosion of security culture. When employees routinely bypass IT-approved tools because they are faster or more convenient, the message is that security is an obstacle, not an enabler. This cultural erosion makes it harder to enforce any security policy.
Blind spots in incident response. When a data breach occurs, the security team needs to trace every channel through which data left the organization. If shadow AI tools are invisible to the security stack, the incident response team cannot determine the full scope of the breach.
How to Discover and Control Shadow AI
Agentless discovery through browser extension telemetry is the most effective way to map shadow AI usage. The extension detects which AI domains the user is interacting with and reports the domain names (not the content) to the central dashboard. Within days, the security team has a complete map of every AI tool in use across the organization.
Once the map is established, each tool can be assigned a policy: allow with DLP scanning, allow in monitor-only mode, or block entirely. The policy is enforced at the browser level, before any data is sent to the AI tool.
The SaaS company that discovered 14 AI tools in use was able to reduce their shadow AI surface to 4 approved tools within two weeks. The 10 unapproved tools were blocked at the browser level. The 4 approved tools were configured with full DLP scanning. In the following month, the detection rate dropped from 347 exposures to 12, all of which were blocked before reaching the AI tool.
Shadow AI is not going away. Employees will continue to adopt AI tools because they make work faster and easier. The goal is not to block all AI tool usage. The goal is to make it visible, governable, and safe.