glmv-pdf-to-web
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using variables derived from external and untrusted sources.
- Evidence: Phase 1 includes
curl -L -o "/tmp/${pdf_stem}.pdf" "$ARGUMENTS", where$ARGUMENTSis a user-supplied URL. - Evidence: Phase 6 includes
python {SKILL_DIR}/scripts/generate_web.py --title "<paper title>", where the title is extracted directly from the PDF content. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF content and uses extracted metadata to influence agent actions, specifically in the construction of command-line arguments.
- Ingestion points: User-provided URLs via
$ARGUMENTSand content extracted from PDF pages (e.g., title, author names, abstract). - Boundary markers: None identified; there are no instructions to sanitize or delimit untrusted strings before use.
- Capability inventory: File system operations (
mkdir), network downloads (curl), and execution of local Python scripts. - Sanitization: Absent; the workflow relies on standard shell quoting which may be insufficient to prevent command injection in complex scenarios.
- [EXTERNAL_DOWNLOADS]: The skill uses
curlto fetch PDF files from arbitrary remote URLs provided at runtime. - Evidence: Phase 1 workflow describes downloading the PDF if the user provides a URL.
Audit Metadata