docx
Warn
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The script
scripts/office/soffice.pycontains an embedded C source string which is written to the system's temporary directory and compiled at runtime usinggcc. The resulting shared library is then injected into the execution environment of LibreOffice usingLD_PRELOADto intercept and redirect system-level socket calls (socket,listen,accept). This is used to enable inter-process communication in restricted environments where standard Unix domain sockets are blocked.\n- [PRIVILEGE_ESCALATION]: TheSKILL.mdfile provides instructions that include usingsudo apt-get installfor managing system-level dependencies such aspandoc,libreoffice, andpoppler-utils. This encourages the agent to request or use administrative privileges, which is a high-risk pattern.\n- [COMMAND_EXECUTION]: The skill makes extensive use of system command execution viasubprocess.run()across several scripts. This includes compiling C code withgcc, running LibreOffice (soffice), converting PDF files withpdftoppm, and performing structural extraction withpandocandgit.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted Word documents as its primary function. Because the skill possesses high-level system capabilities (including command execution and macro usage), it has a vulnerability surface for indirect prompt injection. A maliciously crafted document could potentially influence the agent's behavior during content extraction or analysis.\n - Ingestion points: Document text extracted via
pandocinSKILL.mdand raw XML content fromword/document.xmlread by theDocumentclass inscripts/document.py.\n - Boundary markers: The instructions do not specify strict boundary delimiters or 'ignore instruction' directives when processing external document content.\n
- Capability inventory: Powerful capabilities including file system access, subprocess execution of multiple system tools (
soffice,gcc,pandoc), and execution of LibreOffice macros.\n - Sanitization: While the skill uses
defusedxmlto mitigate XML-specific attacks, there is no evidence of natural language instruction sanitization for processed document content.
Audit Metadata