The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and extract content from untrusted external documents such as scanned PDFs, images, and Office files. This creates a surface for indirect prompt injection where instructions embedded in processed documents could influence the agent's behavior if the output is processed without sufficient sanitization.
Ingestion points: xberg extract command and extract()/extract_batch() functions in the Python, Node.js, and Rust SDKs.
Boundary markers: The documentation does not specify the use of delimiters or warnings to isolate extracted content from system instructions in downstream prompts.
Capability inventory: The skill performs OCR via Tesseract and PaddleOCR, executes layout detection models, generates embeddings via ONNX, and can interface with Vision Language Models (VLMs).
Sanitization: No explicit sanitization or filtering of the extracted text content is mentioned for security purposes.
[PRIVILEGE_ESCALATION]: The instructions in SKILL.md include commands using sudo to install system-level OCR language packs (e.g., sudo apt install tesseract-ocr-deu). This is a legitimate requirement for configuring the underlying OCR engine on Linux systems.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of resources from several sources:
Fetches machine learning models from HuggingFace (e.g., sentence-transformers/all-MiniLM-L6-v2).
Installs the CLI tool via Homebrew, NPM (@xberg-io/xberg-cli), and Cargo.
Pulls official container images from the GitHub Container Registry (ghcr.io/xberg-io/xberg).
[COMMAND_EXECUTION]: The tool invokes external binaries and machine learning runtimes, such as Tesseract OCR and the ONNX Runtime, to perform document analysis and text recognition.

xberg-extracting-with-ocr