The AI You Can't See: Confronting Shadow AI Before It Costs You

Why Employees Turn to Shadow AI

“Every unsanctioned AI tool in your environment is an employee telling you exactly what they need, in the only language they have available,” says Corey Nachreiner, chief security officer at WatchGuard.

They see shadow AI as the latest expression of a familiar pattern: Employees embrace technology that helps them work faster, while organizational approval processes struggle to keep pace. Cloud storage, personal devices and collaboration applications followed similar paths.

AI makes that mismatch especially pronounced. An employee can summarize a lengthy document or generate working code before completing an access request. Treating that behavior solely as a disciplinary issue misses the opportunity to identify a business requirement.

Nachreiner argues that security teams should use unsanctioned adoption to understand which capabilities employees need and where approved offerings fall short.

EXPLORE: How organizations of all sizes are leveraging AI for productivity and collaboration.

Shadow AI Risks: Data Exposure, Access and Response Costs

“Three risks stand out,” says Laura Ellis, senior vice president of artificial intelligence at Arctic Wolf. “First, data exposure: Sensitive and regulated information leaves the organization through prompts and uploads, into tools with no contractual protections.”

The other exposures she identifies are unreviewed access to corporate systems and missing records of what employees used or shared. Together, those gaps leave enterprises uncertain about both the information involved and the permissions an outside service still holds.

Ellis recommends beginning with an inventory. Knowing which tools employees use, who uses them and what data they touch gives leaders a foundation for assessing exposure.

That assessment should distinguish between different business consequences. An application processing public marketing copy presents a different scenario from one receiving customer records or confidential research.

Understanding those differences helps security teams prioritize investigations and estimate potential response costs.

READ MORE: Discover the importance of data governance for a secure AI strategy.

The Agentic Shift: Why Nonhuman Identities Elevate Enterprise Risks

“Agents can inherit excessive permissions and use them to make decisions, call tools and act across systems,” says Jeff Pollard, vice president and principal analyst at Forrester.

That ability changes the stakes, because a tool that can retrieve information, send messages or modify records has opportunities to cause harm beyond disclosing a prompt. Its actions may also diverge from the employee’s intent.

Pollard recommends a “least agency” approach: Grant only the permissions an agent needs and limit how long it retains them. Existing access problems become more consequential when software can exercise excessive privileges automatically.

For cost estimates, he recommends examining specific failure scenarios: What information could escape, what actions could follow, and what investigation, recovery and business interruption would result?

Connecting permissions to those scenarios makes agent risk more concrete for business leaders.

Why Shadow AI Bans Ultimately Fail

Ed Martin, vice president of product management for AI strategy at Sophos, cautions that a ban doesn't stop the use, it just stops the organization from seeing it.

Employees may move to personal phones or home accounts when corporate access disappears, a shift that can leave security teams with less evidence while the underlying demand remains unchanged.

Martin says an approved option that performs worse than a freely available tool invites work-arounds. A workable policy therefore depends on understanding what employees need to accomplish and providing a credible way to accomplish it.

“Every agent needs an identity and owner, the smallest set of permissions that works and an off switch,” Martin says.

DIVE DEEPER: How to combat the security challenges presented by nonhuman identities.

How to Detect Shadow AI Across Network, Endpoint and Cloud Layers

Ellis explains no single control detects all shadow AI, which means organizations must correlate several signals.

Network traffic can reveal connections to AI services, and endpoint monitoring can surface installed applications, browser extensions and locally running models. Identity records show corporate sign-ins and applications authorized to access company information, while cloud monitoring adds visibility into service usage.

However, each perspective leaves gaps, as a local model may never connect to a recognizable AI service, and a personal account may never appear in corporate identity logs.

Ellis recommends combining those observations into an inventory and classifying tools as sanctioned or unknown. Teams can then prioritize services touching sensitive information. 

Meanwhile, discovery must continue as employees adopt new tools and existing applications gain capabilities.

WATCH: Find out how AI is putting additional strain on your network.

Governing the Exposure: AI-Aware DLP and Identity Controls

AI-aware data loss prevention helps identify sensitive information moving into AI systems, including customer records, source code and intellectual property. Controls can address the data transfer that creates exposure.

“The point isn’t to block every interaction,” Nachreiner says. “It is to see the risk clearly enough to apply a proportionate control.”

Strong authentication, least-privilege permissions and conditional access policies help restrict who can reach that information. Those protections need to cover AI agents as well as employees, with ongoing monitoring to detect inappropriate access.

Nachreiner also recommends examining cloud and Software as a Service applications through application discovery and corporate identity monitoring. These provide evidence of AI use that network controls alone may miss, including connections that give outside services access to company resources.

Providing the Alternative: Deploying Enterprise-Sanctioned AI Platforms

Employees need an approved platform that can handle the tasks driving them toward consumer AI. Training and support also matter, particularly when workers are unsure how to apply a tool to their jobs.

“People want four things: good tools, training, community and incentives that don’t introduce problems in their day-to-day,” Martin says.

Leadership support gives employees confidence that the organization wants them to use approved AI. Clear data rules help them understand what they can safely submit, while a quick review process gives them somewhere to turn when an approved tool cannot meet their needs.

Martin recommends a short list of sanctioned tools and explicit guidance about information that must never be entered. Requests for additional capabilities need timely answers; otherwise, employees have an incentive to keep using unapproved alternatives.

RELATED: How managed detection and response services can simplify cybersecurity.

Enablement Over Enforcement: Building Sustainable AI Governance

An AI inventory can help security, IT, legal and business leaders decide which tools to approve, which permissions to restrict and where employees need additional training. Reviewing it regularly also reveals needs that existing platforms have yet to meet.

That requires keeping approval processes responsive as adoption grows. If employees repeatedly seek the same unapproved capability, the organization should evaluate whether to provide it.

If monitoring reveals sensitive data entering a tool, teams need to address the exposure and clarify how employees should handle that task.

“Give employees a path to use AI,” Pollard says. “Move fast to allow them to adopt it, give them the right policies and processes, and encourage them to use AI.”