T05 · Unauthorized Access and Privilege Escalation
- Location
references/examples/gate-approval.md:24- Finding
Caller-Controlled Path Allows Disclosure of Arbitrary Local Files
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a coherent OpenClaw workflow-design helper, but one approval-gate template can read any caller-supplied local file path and the setup guidance uses mutable package installs.
Review before installing. If you use the templates, constrain approval-gate file inputs to an approved workspace/artifact directory, avoid absolute paths and traversal, pin npm/npx package versions, and do not send confidential lesson or business content to llm-task unless that external model use is acceptable.
references/examples/gate-approval.md:24Caller-Controlled Path Allows Disclosure of Arbitrary Local Files
references/lobster-spec.md:177Unpinned Global npm Installation Creates a Supply-Chain Execution Risk
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
.prose/
├── .env
├── runs/
│ └── {YYYYMMDD}-{HHMMSS}-{random}/
│ ├── program.prose
Phần mô tả kích hoạt bao trùm gần như mọi yêu cầu liên quan đến thiết kế, review, sinh file, phân tầng, approval gate, sub-agent, và cả khi người dùng chỉ mô tả mơ hồ bằng lời. Phạm vi quá rộng làm tăng nguy cơ skill bị kích hoạt ngoài ý định, khiến tác nhân đi vào quy trình chuyên biệt, tạo/sửa file workflow trong workspace hoặc đưa ra hướng dẫn kiến trúc không phù hợp với yêu cầu thực tế của người dùng.
Toàn bộ skill được viết và định hướng vận hành bằng tiếng Việt mà không nêu cơ chế chọn ngôn ngữ theo ngữ cảnh người dùng. Điều này có thể làm tác nhân trả lời sai ngôn ngữ mong muốn, hiểu sai yêu cầu bảo mật/kỹ thuật, hoặc tạo artefact workflow và hướng dẫn vận hành không phù hợp với môi trường dùng tiếng Anh của dự án.
The file's headings, notes, and workflow descriptions are entirely in Vietnamese, and there is no indication that users may choose another language or that the skill is intentionally restricted to Vietnamese-speaking contexts. Under the stated policy, forcing a specific language without opt-in is a natural-language policy violation.
The file’s natural-language instructions, labels, and user-facing messages are written entirely in Vietnamese, including invocation examples and approval prompts. Because there is no opt-in, language selection, or stated region-specific scope, this creates a locale policy issue by implicitly forcing a specific language on users.
The skill sends user-provided lesson content via stdin: $lesson_json to openclaw.invoke --tool llm-task, which implies transmission to an external LLM service, but the markdown provides no privacy or data-transfer warning. If users include proprietary curriculum, student information, or unpublished materials, this can cause unintended exposure to third-party processing.
The workflow persists multiple generated artifacts to disk under an output directory without any user-facing disclosure in the skill content. In an agent setting, silent file creation can surprise users, overwrite expected workspace state, or leave behind sensitive derived materials that the user did not realize would be stored.
The entire specification is written in Vietnamese, and there is no indication that users may choose another language or that the locale restriction is required for a region-specific purpose. This creates a natural-language policy issue because the skill documentation effectively imposes a specific language without user opt-in.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.