output-guard-chat-client

Chapter 6 -- OutputGuardChatClient (Critical-3 fix)

Companion code for Generative AI in .NET, Chapter 6 (~line 816) and the Critical-3 fix from review-action-plan.md.

A DelegatingChatClient middleware that scans the model's response text for two common attacks before it reaches the UI:

• Markdown image links pointing to non-HTTP(S) URI schemes (![alt](file:///...), ![alt](data:...)).
• The javascript: URI scheme anywhere in the response.

The Critical-3 fix is the if (response.Text is null) return response; guard. Earlier drafts dereferenced response.Text directly, but the property is null when the assistant returns only tool calls or structured output -- the regex calls would NPE on those responses.

Run it

dotnet run --project samples/ch06-production/output-guard-chat-client

Expected output:

RAW (unsafe):
  Click here for a great offer: [link](javascript:alert(1)) and enjoy this picture: ![cat](file:///etc/passwd)

GUARDED:
  Click here for a great offer: [link](blocked-scheme:alert(1)) and enjoy this picture: [blocked image]

The demo uses a scripted StubChatClient so it runs offline. In production you would chain OutputGuardChatClient into your real provider via ChatClientBuilder.Use(...).

Manuscript reference

• Manuscript/Chapter-06.md, section 6.3.2, "Layer 3 -- Output" (~line 816).
• Manuscript/review-01.md, item C4.

Prerequisites

• .NET 9 SDK