The Agent Skills Directory

[METADATA_POISONING]: The skill includes pre-generated, fake security audit reports intended to mislead reviewers and automated scanning systems.\n
Evidence: The file pdf-extract_audit_result_v2.json contains hardcoded safety claims including "security": "PASS", "gate": "PASS", and "grade": "Production Ready".\n
Evidence: The file report.json explicitly states there are zero findings across all severity levels (high, medium, critical).\n
These files represent a deceptive metadata technique aimed at overriding or influencing the analyzer's own verdict.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted PDF data and interpolates it into the agent's context without safety boundaries.\n
Ingestion points: scripts/extract_pdf.py reads PDF files from user-supplied paths via the --input argument.\n
Boundary markers: Absent. The script generates Markdown output containing extracted text without using delimiters or instructions to prevent the agent from executing embedded commands.\n
Capability inventory: The skill has file system write access for saving Markdown results and image files.\n
Sanitization: Absent. While the text is normalized, it is not sanitized or escaped to neutralize potential prompt injection payloads embedded within PDF documents.

pdf-extract