portable-document-handler
Warn
Audited by Gen Agent Trust Hub on Feb 17, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION] (MEDIUM): The script
scripts/fill_fillable_fields.pyimplements amonkeypatch_pydpf_methodfunction that dynamically overrides theget_inheritedmethod of thepypdf.generic.DictionaryObjectclass at runtime. - Evidence: Code in
scripts/fill_fillable_fields.pyreplaces a library method to work around a specific list-joining bug. - Risk: While intended as a bug fix, dynamic modification of library behavior can lead to unstable state or be used to hide malicious logic by altering standard library calls.
- [PROMPT_INJECTION] (LOW): The
forms.mdfile contains forceful instructional markers designed to override agent decision-making. - Evidence: Phrases such as "CRITICAL: You MUST complete these steps in order. Do not skip ahead" and "Follow the below steps exactly."
- Risk: High-pressure instructions can be used to force the agent to ignore safety guardrails or follow a rigid, potentially unsafe workflow.
- [INDIRECT_PROMPT_INJECTION] (LOW): The skill is designed to ingest and extract content from untrusted PDF files via
pypdf,pdfplumber, and OCR (pytesseract). - Ingestion points:
scripts/extract_form_field_info.py,scripts/check_fillable_fields.py, andscripts/convert_pdf_to_images.pyall process user-provided PDF files. - Boundary markers: The prompt instructions do not specify any delimiters or safety warnings to treat the extracted PDF text as untrusted data.
- Capability inventory: The skill can write to the local filesystem and execute system commands (
qpdf,pdftotext,pdftk). - Sanitization: No sanitization or validation of the extracted PDF text/metadata is performed before it is presented to the agent context.
- [COMMAND_EXECUTION] (SAFE):
SKILL.mddocuments the use of external CLI tools likeqpdf,pdftk, andpdftotextvia subprocess calls. - Evidence: Command examples in
SKILL.mdfor merging, splitting, and decrypting PDFs. - Risk: Standard for the tool's purpose, but requires careful handling of filename arguments to prevent command injection (though the provided scripts use Python libraries primarily).
Audit Metadata