xlsx
Fail
Audited by Gen Agent Trust Hub on Mar 5, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: Runtime code compilation and process injection.
- The file
scripts/office/soffice.pycontains logic to write C source code to a temporary file and compile it into a shared library using thegcccompiler at runtime. - The skill uses the
LD_PRELOADenvironment variable to inject the resulting shared library into thesoffice(LibreOffice) process. This technique is used to intercept and modify standard system calls (such assocket,listen, andaccept), which is a behavior typically associated with unauthorized process manipulation or rootkit functionality. - In
scripts/recalc.py, the skill writes a StarBasic macro (Module1.xba) directly into the LibreOffice user configuration directory (~/.config/libreoffice/...on Linux or~/Library/Application Support/LibreOffice/...on macOS). This modifies the application's global state and persists the macro across different sessions. - Multiple scripts (
pack.py,recalc.py,redlining.py) usesubprocess.runto execute external system tools likesoffice,gcc, andgitwith parameters derived from file paths. - [PROMPT_INJECTION]: Indirect prompt injection surface and deceptive metadata.
- The skill is designed to process untrusted Office documents provided by users. Because it triggers the execution of LibreOffice and macro-based recalculation on these files, it creates a surface for indirect prompt injection where malicious spreadsheet content could exploit the spreadsheet processing logic or the underlying Office suite.
- Mandatory Evidence Chain for Indirect Injection:
- Ingestion points:
scripts/office/unpack.pyandscripts/recalc.pyread and unpack external Office files. - Boundary markers: No specific delimiters or warnings to ignore embedded instructions are provided in the generated code or instructions.
- Capability inventory: The skill has extensive capabilities including file system write access, system command execution (
gcc,git), and application configuration modification. - Sanitization: While
defusedxmlis used for XML parsing, there is no validation or sanitization of spreadsheet formulas or logic before therecalc.pyscript triggers a full recalculation via LibreOffice. - Deceptive Metadata: There is a significant discrepancy between the author context provided (
weihuizhong) and the internal file metadata (e.g.,LICENSE.txtandSKILL.md), which claims the materials are proprietary toAnthropic, PBC. This misleading information can lead to a misjudgment of the skill origin and its associated risks.
Recommendations
- AI detected serious security threats
Audit Metadata