Skip to main content
ShadowAI Defense GitHub
← All detections
DET-ZS-003 Zscaler Network config Severity: Medium

Zscaler — Cloud browser isolation for AI services

Zscaler Browser Isolation policy that renders AI service sessions in an isolated cloud container, providing full visibility into user interactions while preventing direct data transfer to AI platforms.

Rule

# Zscaler — Cloud Browser Isolation (CBI) for AI Services
# Navigation: ZIA Admin Portal > Policy > Cloud Browser Isolation
#
# Isolation Profile: AI Service Isolation
#
# Profile settings:
#   - Name: Isolate AI Services
#   - Isolation type: Isolation with controls
#
# Data controls (within isolated session):
#   - Copy to clipboard: Block (prevent copy from corporate apps into isolated AI session)
#   - Paste from clipboard: Block (prevent paste into AI chat/upload fields)
#   - File upload: Block (prevent file upload to AI services)
#   - File download: Block (prevent downloading AI-generated content with embedded data)
#   - Print: Allow (for audit trail — user can screenshot but effort is logged)
#   - Keyboard input: Allow (manual typing permitted — limits bulk data transfer)
#
# Alternatively, for monitoring without blocking:
#   - Copy/Paste: Allow but log
#   - File upload: Allow but log + DLP inspect
#   - This mode provides full visibility while maintaining productivity
#
# URL Filtering Rule (pairs with isolation profile):
#   - Name: Isolate AI/ML Applications
#   - URL Category: 'AI/ML Applications'
#   - Action: Isolate
#   - Isolation profile: AI Service Isolation (defined above)
#   - Users: All users (or phased by department)
#   - Exceptions: Sanctioned enterprise AI with dedicated allow rule above
#
# Logging:
#   - All isolated session activity logged including keystrokes-to-AI and paste attempts
#   - Session recordings available for incident investigation (if licensed)
#   - Forward logs to SIEM via NSS for correlation

How it works

Zscaler's Cloud Browser Isolation (CBI) renders web sessions in a remote container, streaming only pixels to the user's browser. When applied to AI services, this creates a controlled environment where users can interact with AI tools but cannot transfer bulk data via copy-paste, file upload, or download. This approach is particularly effective for organizations that want to allow AI experimentation while preventing data exfiltration — it provides a middle ground between full block and full allow. The isolation layer also generates detailed session logs that are unavailable with standard proxy logging, capturing actual interaction patterns.

Required data sources

  • Zscaler Internet Access (ZIA) with Cloud Browser Isolation (CBI) add-on
  • Zscaler Client Connector deployed on managed endpoints

Prerequisites

  • Zscaler Internet Access with Cloud Browser Isolation (CBI) licensed
  • Zscaler Client Connector deployed to managed endpoints
  • SSL/TLS inspection enabled for AI service domains
  • CBI isolation profiles configured with appropriate data controls
  • User communications explaining the isolated-browser experience for AI services

Expected volume

Isolation session logs: 50-200 isolated sessions per day per 10k users. Blocked paste/upload attempts: 10-40 per day initially, dropping as users adjust behavior.

False-positive guidance

Browser isolation may degrade the user experience for legitimate AI tool usage — some AI services have complex UIs that render differently in isolation. Test top AI services in isolation before broad rollout. Some browser extensions may not function within isolated sessions — coordinate with users who rely on accessibility extensions.

Tuning steps

  1. Start with monitoring mode (allow all actions but log) for 2 weeks to understand usage patterns.
  2. Enable clipboard blocking for unsanctioned AI services while keeping sanctioned services unrestricted.
  3. Review session logs weekly to identify high-risk users transferring sensitive data.
  4. Test user experience in isolation for each major AI service before enforcing.
  5. Consider allowing file upload with DLP inspection as a middle ground for departments with legitimate AI use cases.

Framework mappings

  • NIST AI RMF: GOVERN-4.1, MANAGE-2.2
  • ISO/IEC 42001: 8.2, 8.3
  • NIST CSF: PR.AC-3, PR.DS-5, DE.CM-7

Related AI Controls Catalog entries