pptx
Warn
Audited by Gen Agent Trust Hub on Mar 6, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill performs dynamic code generation and system-level injection to handle process communication in restricted environments.\n
- Evidence: The file
scripts/office/soffice.pyembeds C source code that is written to disk, compiled viagcc, and then loaded into thesofficeprocess usingLD_PRELOAD. This intercepts and redirects standard socket calls.\n- [PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection (Category 8) through its visual inspection features.\n - Ingestion points: Processes untrusted text and layouts from
.pptxfiles usingscripts/thumbnail.pyandmarkitdown.\n - Boundary markers: The instructions for the Visual QA subagent in
SKILL.mdlack delimiters to separate user data from instructions.\n - Capability inventory: The skill can execute arbitrary commands (
subprocess) and manipulate the file system (unlink,rmtree) via scripts likescripts/clean.pyandscripts/office/pack.py.\n - Sanitization: No sanitization is applied to the data extracted from presentation files before it is used in subagent prompts.
Audit Metadata