Skip to content

Response-side inspection (outbound filtering) #31

@temp-noob

Description

@temp-noob

Why

Every existing OSS MCP firewall only inspects requests (tool call arguments). None inspect responses (data returned by the MCP server). This is an explicitly acknowledged gap — the mcpwall DEV article 'What mcpwall Does and Doesn't Protect Against' calls this out as a missing capability.

A compromised or malicious MCP server can leak secrets, PII, or encoded exfiltration payloads in its responses, and no guardrail catches it.

What

Inspect MCP server responses before forwarding back to the agent client.

Acceptance Criteria

  • Scan responses for secret patterns (API keys, tokens)
  • Scan responses for PII (configurable)
  • Detect base64-encoded suspicious content in responses
  • Policy YAML section: response_rules with configurable patterns
  • Action: block response, redact matches, or warn
  • Tests: compromised server returning secrets in response

Competitive advantage

This would make IntentGuard the first OSS MCP guardrail with response-side inspection. Unique differentiator.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions