parse-document

Pass

Audited by Gen Agent Trust Hub on May 13, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill facilitates the extraction of text from external documents, which creates an indirect prompt injection surface. Maliciously crafted content within a PDF or image could attempt to override the agent's instructions once the text is parsed and read into the context.
  • Ingestion points: Local files are processed through the document_parse tool in SKILL.md.
  • Boundary markers: The instructions do not provide specific delimiters or warnings for the agent to treat the parsed content as untrusted data.
  • Capability inventory: The agent uses the read tool to inspect the parsed output and performs downstream tasks like summarization.
  • Sanitization: No sanitization or filtering of the extracted text is mentioned in the skill instructions.
  • [DATA_EXFILTRATION]: The support for the ocrServerUrl parameter allows the agent to send document data to an external network endpoint. While the instructions advise using this only when requested by the user, it remains a potential vector for exfiltrating sensitive document content to an unauthorized or untrusted OCR service.
Audit Metadata
Risk Level
SAFE
Analyzed
May 13, 2026, 09:01 PM
Security Audit — agent-trust-hub — parse-document