glmv-pdf-to-ppt
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads PDF files from user-supplied URLs using the
curlcommand to a local temporary directory. - [COMMAND_EXECUTION]: The workflow relies on executing several local shell commands and Python scripts, including
mkdir,curl, and internal skill scripts for PDF conversion, image cropping, and slide generation. User-provided input ($ARGUMENTS) is interpolated into shell commands in Phase 1, which represents a potential injection surface if the environment does not provide adequate sanitization. - [DYNAMIC_EXECUTION]: In Phase 5, the skill dynamically constructs and executes a Python code snippet (
python3 -c ...) to measure image dimensions and calculate aspect ratios for slide layout planning. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external PDF files and URLs. A malicious document could contain instructions designed to influence the agent's behavior during the outline planning or slide generation phases.
- Ingestion points: PDF images processed via the agent's vision capabilities and PDF files downloaded from external URLs.
- Boundary markers: No explicit instructions or delimiters are used to warn the agent or subagents about potential instructions embedded within the processed document content.
- Capability inventory: The skill can execute shell commands, run Python scripts, and write multiple file types (JSON, Markdown, HTML) to the workspace.
- Sanitization: The skill does not specify any sanitization or filtering of the data extracted from the PDF before it is used to generate the presentation outline or slide content.
Audit Metadata