Skip to main content
ShadowAI Defense GitHub
← All detections
DET-NET-001 Generic Network Network config Severity: Medium

Generic SWG / proxy — block consumer AI domains

Stack-agnostic secure web gateway / proxy configuration to block consumer AI domains for unauthorized users.

Rule

# Stack-agnostic SWG / Proxy / Firewall policy
#
# Block category: AI / Generative AI (where the SWG vendor provides this category)
# Custom URL list source: /watchlists/ConsumerLLMDomains.csv from this repo
#
# Action:
#   - Block with custom block page
#   - Block page references the AUP and links to /approved-tools (internal portal)
#   - Allow-list users in approved AI pilot groups (sync from Entra ID / Okta group)
#
# Logging:
#   - All allowed and blocked AI-category traffic logged with user, source, destination, timestamp
#   - Logs forwarded to central SIEM for correlation with DET-MS-SEN-001 / equivalents
#
# Supported vendors with native 'Generative AI' category (verify currency):
#   - Zscaler (URL Category: Generative AI)
#   - Netskope (App Tag: Generative AI)
#   - Palo Alto Prisma Access (URL Filtering Category: ai-and-ml-applications)
#   - Cisco Umbrella (Generative AI category)
#   - Forcepoint Web Security (AI tools category)
#   - Fortinet FortiGate (AI Chatbot URL Filter category)

How it works

Network-layer control for any organization with an SWG/proxy. Most modern SWGs ship a Generative AI URL category — use it directly. For organizations without that category, populate a custom URL list from the watchlist file.

Required data sources

  • Any modern SWG / proxy / next-gen firewall with URL filtering

Prerequisites

  • SWG / proxy / firewall enforcing for managed endpoints
  • Identity integration so policy can apply per group (or by source IP / VLAN as fallback)

Expected volume

Each blocked attempt generates one log; expect a sharp peak day 1 dropping to a stable baseline after AUP + approved-tools comms.

False-positive guidance

Confirm category curation freshness with the vendor; over-block on legitimate non-AI sites occasionally seen with new categorizations.

Tuning steps

  1. Sync allow-list weekly from Entra ID / Okta sanctioned AI users group.
  2. Refresh custom URL list monthly from the published watchlist.
  3. Review block-page click-throughs to /approved-tools as a comms-effectiveness metric.

Framework mappings

  • NIST AI RMF: GOVERN-4.1
  • ISO/IEC 42001: 8.2
  • NIST CSF: PR.AC-3, DE.CM-1

Related AI Controls Catalog entries