kill-argument
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document content (LaTeX, BibTeX, and PDF files) provided in the target directory. This content is passed to model-based review tools (Codex MCP) without sanitization or boundary markers, allowing malicious instructions embedded in a paper to potentially influence the agent's behavior.
- Ingestion points: All LaTeX source files, bib files, and PDFs found within the directory provided via
$ARGUMENTS(SKILL.md, Step 1). - Boundary markers: Absent. The content is interpolated into prompts without delimiters or instructions to ignore embedded commands.
- Capability inventory: The skill utilizes
Bash,Read,Write, andEdittools, as well as multiple MCP review tools for analysis and adjudication. - Sanitization: No sanitization or filtering is performed on the content of the ingested files before they are sent to the model.
- [COMMAND_EXECUTION]: The skill uses
Bash(*)to execute file discovery commands. The use of$ARGUMENTSto define the working directory (PAPER_DIR="$ARGUMENTS") introduces a potential command injection surface if the input is not properly sanitized by the underlying platform before being executed in a shell environment. - Evidence: The script in Step 1 of the workflow directly assigns the user-provided
$ARGUMENTSto a variable and uses it in acdcommand.
Audit Metadata