pdf

Warn

Audited by Gen Agent Trust Hub on Feb 15, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • COMMAND_EXECUTION (LOW): The skill invokes several local Python scripts (e.g., check_fillable_fields.py, convert_pdf_to_images.py) using the system shell and specific environment variables like PYTHONPATH=gateway/tools. This is a standard functional pattern but assumes the local tools directory is secure and not writable by the agent itself.
  • PROMPT_INJECTION (MEDIUM): This skill exhibits a significant Indirect Prompt Injection (Category 8) attack surface because its core function is to ingest and act upon untrusted PDF data.
  • Ingestion points: Untrusted data enters the agent's context through file paths like input.pdf processed by tools such as pdfplumber and pypdf in 'Phase 2-A'.
  • Boundary markers: The skill does not define clear delimiters or instruction-isolation markers (e.g., XML tags or headers) to prevent the agent from mistaking text inside a PDF for system-level instructions.
  • Capability inventory: The agent has the capability to write files, execute subprocesses, and delegate tasks to subagents based on the content parsed from these PDFs.
  • Sanitization: There is no evidence of sanitization or filtering logic to strip potential malicious instructions from the extracted PDF text before it is used to influence the workflow.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 15, 2026, 07:42 AM