docx

Warn

Audited by Gen Agent Trust Hub on Apr 18, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill performs dynamic execution by compiling and injecting code at runtime to facilitate process communication.
  • In scripts/office/soffice.py, the skill contains an embedded C source shim designed to bypass environment-specific socket restrictions when running office software.
  • It invokes gcc at runtime to compile this source into a shared object file within the system's temporary directory.
  • The resulting library is then injected into the LibreOffice (soffice) process using the LD_PRELOAD environment variable. While this supports the primary functionality of document conversion, runtime compilation and process injection are high-privilege patterns.
  • [COMMAND_EXECUTION]: The skill relies on several external binaries and provides a wide range of system capabilities to the agent.
  • It uses subprocess.run to execute soffice for file conversion, pandoc for document parsing, pdftoppm for image generation, and git for version comparison.
  • These tools are used appropriately for the skill's purpose but represent a significant capability set available to the agent for interacting with the host system.
  • [PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection (Category 8) due to its document processing capabilities.
  • Ingestion points: Untrusted data enters the agent context when unpacking and parsing XML files from Word documents in scripts/office/unpack.py and scripts/office/validators/docx.py.
  • Boundary markers: Explicit delimiters or specific instructions to ignore embedded commands within the document XML are not implemented in the prompts.
  • Capability inventory: The skill has the ability to execute shell commands (subprocess.run in scripts/office/soffice.py, scripts/accept_changes.py, and scripts/office/validators/redlining.py) and perform file system operations across its included scripts.
  • Sanitization: There is no evidence of content sanitization or instruction filtering for the text extracted from Word documents before it is processed by the agent.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 18, 2026, 02:18 AM