AI 治理

AI data security is the practice of discovering, classifying, and controlling access to the sensitive data that Copilot, LLMs, and AI agents can reach, then monitoring how that data is used once AI systems are in production. It differs from AI governance, which focuses on model behavior, policy, and compliance. Netwrix approaches AI data security by exposing sensitive and shadow data, revealing risky access, and reducing exposure before AI amplifies it.

AI governance covers the policies, model oversight, and compliance controls that govern how an AI system behaves. AI data security covers a different question: what sensitive data can that AI system actually reach, and is that access appropriate? An organization can have a mature AI governance program and still expose sensitive data if the underlying access and identity controls are not addressed. Netwrix's 2026 Identity and Data Security Convergence Survey found that AI and automation increased the number of identities with access to sensitive data at 58.8% of organizations, which is a data security problem, not a policy problem.

Copilot can surface any file or record that a user's existing permissions already allow them to reach, including overexposed files, stale permissions, and content in shared drives. If those permissions were never tightened for a world where AI can search and summarize everything at once, Copilot can expose sensitive data that was previously just difficult to find manually. Only 21.1% of organizations report complete visibility into the sensitive data their AI tools are using, per the 2026 Netwrix Identity and Data Security Convergence Survey. Assessing and tightening data access before a Copilot rollout is the most direct way to close that gap.

A written AI policy does not, by itself, reduce risk. An effective program needs an inventory of every AI agent, copilot, and automated identity in the environment, a map of what sensitive data each one can reach, enforced least privilege, and continuous monitoring rather than periodic reviews. Only 11% of organizations describe themselves as "Ready" on AI security readiness, per the 2026 Netwrix Identity and Data Security Convergence Survey, and 17.2% report no formal AI data inventory or governance process at all.

AI agents and Copilot inherit whatever access already exists in the environment: file shares, SharePoint sites, service account permissions, and API scopes. Mapping this requires tools that can enumerate effective permissions across hybrid environments and flag which identities, human or non-human, can reach sensitive data. Netwrix Access Analyzer and 1Secure provide this visibility, including specifically for Copilot and AI agent access.

Shadow AI is the use of AI tools, models, or copilots that IT and security teams have not approved, reviewed, or included in governance processes. Employees pasting sensitive data into public chatbots, or teams connecting unsanctioned AI tools to internal systems, are both common forms of shadow AI. It is a growing blind spot: many organizations lack full visibility into which AI tools employees are actually using day to day.

Both, but they require different controls. Governing what users do with AI means acceptable-use policies: which AI tools are approved, and what employees are allowed to do with them. Governing what AI can access means controlling the sensitive data and permissions that Copilot, LLMs, and AI agents can reach once they are deployed, regardless of how carefully employees follow policy. A user can follow every rule and still have Copilot surface data they were never supposed to see, because the exposure was already baked into stale permissions. Most AI governance conversations start with policy and skip the access question. Netwrix focuses on the access side: discovering, classifying, and controlling what AI can actually reach.

Blocking or monitoring what employees type into AI tools like ChatGPT, Gemini, or Copilot chat requires data loss prevention at the endpoint, not just an acceptable-use policy. Netwrix Endpoint Protector can detect and block sensitive data such as PII, PHI, source code, and financial data from being pasted or uploaded into unsanctioned AI tools, and can tag content so policies still trigger even when a file does not obviously match a sensitive pattern.