grad-elm
Pass
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The file
examples/sample_scenario.mdcontains embedded instruction-like text that appears to be a leaked transcript from an AI agent's internal file-writing operation. - Ingestion points: The text is located in
examples/sample_scenario.md, which is intended to be processed as data by an agent using this skill. - Boundary markers: The file lacks explicit delimiters or instructions telling the agent to ignore the meta-instructions embedded in the prose.
- Capability inventory: The skill itself does not request dangerous tools, but the presence of command-like text ('讓我直接寫入目標檔案', 'Writing it now') targeting a specific file path could confuse downstream agents that possess file-writing capabilities.
- Sanitization: No sanitization or escaping is applied to these apparent traces of internal agent operations.
Audit Metadata