The Agent Skills Directory

[SAFE]: The skill code is a defensive utility designed to protect agents from malicious instructions embedded in external content.
[EXTERNAL_DOWNLOADS]: Fetches pre-trained classification models from Hugging Face organizations (meta-llama, protectai) and standard security references (MITRE, OWASP).
[COMMAND_EXECUTION]: Relies on the external Tesseract OCR binary for image text extraction, which is consistent with the skill's analytical purpose.
[PROMPT_INJECTION]: The skill processes untrusted external data artifacts as its core function.
Ingestion points: extract_html, extract_pdf, and extract_image functions in scripts/agent.py.
Boundary markers: Normalization logic is used to strip zero-width and Unicode tag characters, though specific prompt delimiters for the final LLM context are not defined in the skill scripts.
Capability inventory: The agent script is limited to reading input files, performing local analysis, and writing JSON telemetry results.
Sanitization: Employs a multi-layered approach including regex heuristics, normalization (Base64/ROT13 decoding), and transformer-based classifier models.

detecting-indirect-prompt-injection