Skip to content

Tool description poisoning / rug-pull detection #32

@temp-noob

Description

@temp-noob

Why

MCP servers can change their tool descriptions, argument schemas, or capabilities after being initially trusted. This 'rug-pull' attack lets a trusted tool silently change behavior — e.g., a read_file tool that starts silently exfiltrating data to an external endpoint.

No existing OSS guardrail tool detects this.

What

Snapshot MCP tool metadata on first proxy connection, detect and alert on changes.

Acceptance Criteria

  • On first connection to an MCP server, snapshot the tools/list response (tool names, descriptions, argument schemas)
  • Store snapshot in a local file (.intent-guard/tool-snapshots/<server-hash>.json)
  • On subsequent connections, compare current tools/list with snapshot
  • Alert (audit log + stderr warning) if any tool has changed description, schema, or new tools appeared
  • Configurable: --detect-tool-changes flag
  • Policy option: block changed tools until re-approved, or warn-only
  • Tests: simulated rug-pull scenario

Competitive advantage

Novel capability — no OSS or commercial tool currently detects MCP tool description poisoning.

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