webmcp
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes a system for handling external data through tool parameters, identifying an indirect prompt injection attack surface.\n
- Ingestion points: Tool parameters are passed to the
executecallback as defined inAGENTS.mdandrules/tool-definition.md.\n - Boundary markers: The skill encourages the use of
agent.requestUserInteraction()(documented inrules/user-interaction.md) to establish a human-in-the-loop confirmation boundary.\n - Capability inventory: Tools are intended to interact with browser-based application logic, including potentially sensitive actions like purchases or account deletions.\n
- Sanitization:
rules/security.mdprovides explicit instructions to validate all agent-provided parameters and warns about 'Tool poisoning' and 'Output injection'.\n- [EXTERNAL_DOWNLOADS]: The skill identifies external resources for WebMCP implementation.\n - Evidence:
AGENTS.mdreferences the@mcp-b/globalpolyfill and thehttps://github.com/WebMCP-orgGitHub organization.\n - Context: These references are provided to assist developers in implementing the proposed W3C WebMCP standard.
Audit Metadata