T05 · Unauthorized Access and Privilege Escalation
- Location
SKILL.md:10- Finding
Mandatory Access to Private Vault and Persistent Agent-Memory Files
- Content
View full analysis
Vulnerability Details
File Location:
SKILL.md:10-14
Additional Location:SKILL.md:57
Vulnerability Type: Excessive access to private project context and persistent Agent memory
Risk Level: MediumVulnerable Code Snippet:
markdown 1. `~/Documents/ntriq-vault/wiki/decisions/2026-04-19-android-cli-채택.md` — 도입 결정/스택 2. `~/Documents/ntriq-vault/wiki/tools/android-cli.md` — CLI 사용법·커맨드 레퍼런스 3. `~/Documents/ntriq-vault/wiki/decisions/2026-04-18-faktum-news-첫-사업-확정.md` — Faktum 사업 컨텍스트 4. `~/Documents/ntriq-vault/wiki/system/faktum-news-current-state.md` — 현재 웹 스택/DB/API 스냅샷 5. `~/.claude/projects/-Users-youareplan/memory/project-faktum-news.md` — 프로젝트 메모리 + 모바일 섹션The workflow also requires access to application source outside the Skill package:
markdown 5. Faktum 웹 API 계약 Read (`~/Documents/Projects/faktum-news/app/api/`)Technical Analysis
The Skill requires the Agent to read files outside both the Skill package and the prospective Android project. These files include a synchronized private vault, internal business and architecture records, persistent Claude project memory, and application API source.
This violates least-privilege principles because access is mandatory and is not constrained by explicit user approval, file-content filtering, or a minimum-data requirement. Persistent Agent-memory files may contain historical instructions, user-specific information, or unrelated project context. Loading that information into the active session expands the amount of sensitive data available to all subsequent instructions and tools.
There is no evidence that the Skill transmits the accessed information to an external party. The risk is therefore unauthorized or excessive local information access rather than confirmed data exfiltration.
Attack Path
- A user invokes the Skill through one of its configured Android or Faktum-related trigger phrases.
- The Agent follow ...[truncated 1098 chars]
- Remediation
View remediation
Remediation Suggestions
- Remove the requirement to read the persistent Claude memory file by default.
- Ask for explicit user consent before accessing any path outside the active project.
- Replace broad mandatory reads with a list of optional, task-specific context files.
- Request a minimal API schema or selected endpoint files instead of reading the entire API directory.
- Use project-relative paths or user-approved configuration rather than hardcoded, user-specific absolute paths.
- State which information is required from each file and prohibit use of unrelated content.
- Add a pre-access summary that lists every external path and allows the user to approve or reject each one.
- Ensure sensitive content loaded from these files is not copied into generated source, logs, or external requests unless separately authorized.
