ljg-qa
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external content from URLs (via WebFetch) and PDF files (via Read tool) to extract information. This creates a surface for indirect prompt injection where malicious instructions embedded in a processed document could attempt to influence the agent's output or logic.
- Ingestion points:
Workflows/Extract.md(Step 1) details fetching content from URLs, arXiv links, and local PDFs. - Boundary markers: The instructions do not define specific delimiters (e.g., XML tags or clear separators) to isolate external content from the skill's operational instructions.
- Capability inventory: The skill utilizes
WebFetch,Read,WebSearch, file writing to~/Documents/notes/, and shell execution for notifications. - Sanitization: There is no evidence of sanitization or filtering applied to the external content before it is processed by the LLM.
- [COMMAND_EXECUTION]: The skill instructions include shell commands intended to be executed during the workflow. Specifically, it uses
curlto send a POST request tohttp://localhost:31337/notifyand thedatecommand for generating timestamps. While targetinglocalhostis generally considered low-risk for exfiltration, this establishes a pattern of side-effect execution on the local environment.
Audit Metadata