Li-Material-Mining
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill references absolute file paths within the
/Users/liwanling/directory. Hardcoding local user-specific paths exposes the directory structure and user account naming convention of the execution or development environment. - Evidence: Multiple paths in the 'Internal Method Box' (内部方法箱) table, such as
/Users/liwanling/.agents/skills/and/Users/liwanling/Downloads/李同学/. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted narrative input from users (daily stories and dialogue) but does not implement boundary markers to isolate this data from its operational instructions.
- Ingestion points: User replies to the agent's interview-style questions regarding their day (SKILL.md).
- Boundary markers: None present in the prompt templates.
- Capability inventory: The agent is instructed to use a
Readtool to access local methodology files based on the processed input (SKILL.md). - Sanitization: No evidence of input escaping or validation is present.
Audit Metadata