Module 7 — AI Controls (Safe AI Usage)

Goal: See which AI tools your people use, stop sensitive data from being pasted into them, and protect any AI apps you build yourself.

🧭 The golden rule for AI: govern, don't ban. If you hard-block ChatGPT, people just use it on their phones — and you lose all visibility. The goal is to allow AI safely: see it, control risky actions, and stop sensitive data going in.

Cloudflare's AI Security Suite — the big picture

This module (and Modules 5d, 6, 7b, 7c) together make up Cloudflare's AI Security Suite: secure and govern AI everywhere innovation happens, across the whole AI lifecycle — so you can scale AI adoption without sacrificing security. It's worth seeing how the pieces fit before diving in.

Four outcomes the suite delivers:

Four use cases across the AI lifecycle:

📎 One capability this onboarding doesn't set up in depth: CASB / AI-SPM. Cloudflare's CASB adds AI Security Posture Management — it scans your sanctioned SaaS and AI tools for misconfigurations (over-shared data, risky integrations, weak settings) and helps you fix them. It complements the in-traffic controls in this suite with at-rest posture checks on the AI tools you've approved. If you've licensed CASB, run its AI-SPM scans alongside these modules. (Reference: cloudflare.com/solutions/ai-security.)

We'll do: (A) discover what's used → (B) allow-but-restrict → (C) protect prompts with DLP → (D) protect AI you build → (E) govern AI agents.

Part A — Discover which AI tools are in use

1. 👉 Zero Trust → Gateway → Analytics → Shadow IT Discovery (or My Team → Shadow IT Discovery).
2. 👉 Filter the application type to Generative AI.
3. 📺 What you'll see: A list of AI apps your users have accessed (ChatGPT, Gemini, Claude, Perplexity, Copilot…), with how many users and how much traffic.
4. 👉 For each app, set a status: mark sanctioned tools Approved and others Unapproved.

✅ Checkpoint: You now have an evidence-based picture of AI usage — the basis for your AI acceptable-use policy.

🔎 The full discover→review→enforce workflow (for all SaaS, not just AI) plus the AI-adoption journey is in Module 5d — Shadow IT & AI Security Adoption.

Part B — Allow AI, but restrict risky actions

Instead of blocking an AI app, allow it while blocking the risky parts (like file uploads).

1. 👉 Zero Trust → Gateway → Firewall Policies → HTTP → Add a policy.
2. ⌨️ Name: AI - allow with guardrails.
3. 👉 Rule: Selector Application → in → select your AI apps (e.g. ChatGPT, Google Gemini, Claude, Perplexity).
4. 👉 Expand Application granular controls and block specific actions, e.g.:
   • Block Upload file
   • (optionally) Block Share / create actions
5. 👉 Action: Allow (with the granular blocks applied).
6. 👉 Click Create policy.

📺 What this does: People can chat with the AI, but can't upload your documents into it.

✅ Checkpoint: On a pilot device, you can use ChatGPT normally, but attempting to upload a file is blocked.

Part C — Protect prompts with DLP (Enterprise)

This scans what users type into AI tools and blocks prompts containing sensitive data.

Step C1 — Turn on AI prompt detection

1. 👉 Zero Trust → DLP → Detection entries (or open a DLP profile) → find AI prompt topics.
2. 👉 Enable the topics you care about:
   • Content topics: PII, Source Code, Credentials & Secrets, Financial Information, Customer Data
   • Intent topics: jailbreak attempts, requests for malicious code, attempts to extract personal data
3. 💡 Or just select a ready-made profile like "AI Prompt: PII" or "AI Prompt: AI Security."

Step C2 — Enforce it (monitor first!)

1. 👉 Gateway → HTTP → Add a policy named AI prompts - MONITOR.
2. 👉 Rule: Selector DLP Profile → in → your AI Prompt profile. Action Allow (logs only).
3. 👉 Run it for a few days; review Gateway → Logs.
4. 👉 Then add a higher-priority policy AI prompts - BLOCK with Action: Block for the high-confidence cases (e.g. prompts containing Credentials & Secrets).

✅ Checkpoint: Pasting a fake secret/API key into an AI chat is detected (and blocked by the block policy), visible in logs.

Part D — Protect AI apps you build

Two different tools, depending on how AI is used:

D1 — For programmatic / API AI (no device client needed): DLP for AI Gateway

If your developers call AI models through Cloudflare AI Gateway:

1. 👉 Cloudflare dashboard → AI → AI Gateway → select your gateway.
2. 👉 Open Features → DLP → Set up.
3. 👉 Attach DLP profiles to scan the requests and responses to/from AI providers.

💡 This works without TLS decryption or the device client — it inspects traffic at the API layer. Great for governing app-to-AI calls.

🚪 AI Gateway does much more than DLP — authentication, guardrails, cost/rate limits, caching, and logging for every AI call your apps make. Full walkthrough: Module 7c — AI Gateway: Protect & Implement.

D2 — For AI apps you expose to users: AI Security for Apps (WAF)

If you host an AI app/chatbot that others reach over the web:

1. 👉 Go to the zone (website) hosting your AI app → Security → Settings (or WAF → Detections).
2. 👉 Enable AI Security for Apps detections — prompt injection and unsafe topics.
3. 👉 Create a WAF custom rule that acts (Block/Log) when those detections fire.

💡 How they differ: AI Security for Apps stops attacks on your model (prompt injection); DLP stops sensitive data moving. Use both on AI traffic.

🕷️ The other inbound threat: AI crawlers. Beyond attacks, AI bots crawl your public content to train models — now the majority of internet traffic. To see, allow/block, or even charge them, see Module 7d — The Agentic Internet: AI Crawler & Bot Control.

Part E — Govern AI agents (MCP server portals) — optional/advanced

If you expose tools to AI agents via the Model Context Protocol (MCP):

1. 👉 Zero Trust → Access controls → AI controls → MCP server portals.
2. 👉 Add a portal, publish the tools/prompts you want to expose, and put an Access policy in front so only authorized agents/users can use them.

🔐 This is a big topic on its own. For the full walkthrough — securing individual MCP servers, building an MCP portal, Managed OAuth for AI clients, service tokens for bots, and request-level logging + DLP — see the companion page: Module 7b — Secure AI & MCP with Access (MCP Server Portals).

✅ Module 7 complete!

You now have:

• ✅ Visibility into which AI tools are used (Shadow IT)
• ✅ An "allow but restrict" policy for AI apps
• ✅ (Enterprise) Prompt-level DLP detecting/blocking sensitive prompts
• ✅ Protection for AI you build (API layer and/or web layer)
• ✅ (Optional) Governance for AI agents via MCP portals

Quick troubleshooting

👉 Next: Module 8 — Cloudflare WAN

Connect whole offices and data centers to Cloudflare. (Optional; requires Enterprise network add-on.)