WeaknessASC-2026-0002Major weakness
Tool misuse through over-broad MCP tools
HighMitigated
Agents call legitimate tools in harmful ways: shell, file-write, email or payment tools with no scoping or argument checks.
How to detect it
- MCP tool list exposes shell execution, arbitrary file write or unrestricted HTTP
- Tool arguments accept free-form commands, paths or URLs
How to resolve it
- Allow-list tools per agent and per environment
- Constrain arguments: path roots, domain allow-lists, enums instead of free text
- Split read and write tools; put write tools behind human approval
How to investigate
- Pull tool call logs for the session and look for arguments outside the expected scope
- Correlate each unexpected call with the content the agent read just before it
- Check what the tool actually did (files changed, requests sent) and roll back