Skip to content

Input hardening: path confinement policy + resource limits for file-taking tools (follow-up finding from PsychQuant/macdoc#134 verify) #166

Description

@kiki830621

Problem

6-AI verify(R1 findings #2/#9/#14,DA 已降級 severity 但值得追蹤):全部 245 tools 沿用「caller 提供任意路徑即讀寫」的 local-MCP 信任模型,無根目錄限制、無覆寫保護、無檔案大小/zip-entry 上限。script-pipeline tools 與既有 surface 一致,但 server-wide 的 input-hardening 政策值得一次設計:可讀寫根目錄 allowlist(opt-in config)、既有檔案覆寫確認參數、讀取上限(大檔/zip-bomb)。

Source

Refs PsychQuant/macdoc#134


Current Status

Phase: diagnosed
Last updated: 2026-07-16 by idd-diagnose

Key Decisions

Scope Changes

  • (none)

Blocking

  • (none)

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions