.Net: samples: add prompt-injection + tool-call hardening examples (filters)

[aeris-systems]

Draft: PR description (Semantic Kernel sample: prompt audit + tool gating)

Summary

Adds a small sample showing how to build a pre-validation sensor pipeline in Semantic Kernel using:

• IPromptRenderFilter for prompt audit + preflight scanning
• IAutoFunctionInvocationFilter for tool gating (fail-closed for high-risk operations)

Motivation

Teams are increasingly running SK agents against untrusted sources (web pages, PDFs, email, issue comments). The dominant practical risk is tool-output injection (malicious content steering the agent to exfiltrate secrets or invoke dangerous tools). The sample demonstrates a pragmatic mitigation pattern without changing SK core.

What’s included

• Policy tags (allow / require-confirmation / block)
• Prompt render audit log
• Function invocation guard:
  • blocks obviously unsafe requests
  • requires confirmation for destructive/high-privilege tools
  • fails closed when the “sensor” is unavailable
• Optional scanner hook (HTTP call) you can replace with your own service

Notes

• Sample-only; no breaking changes.
• Intentionally minimal so users can copy/paste into production.

Testing

• dotnet test for the sample project
• Manual run: shows audit log + blocks a simulated injection

[aeris-systems]

@microsoft-github-policy-service agree company="Aeris Systems"