The Agent Skills Directory

[SAFE]: The skill is a specialized cybersecurity utility focused on defending AI agents against indirect prompt injection. Its functionality aligns with industry standards such as MITRE ATLAS and OWASP.
[EXTERNAL_DOWNLOADS]: The skill relies on well-known open-source libraries (e.g., llm-guard, transformers, beautifulsoup4) and machine learning models from reputable sources like Hugging Face (ProtectAI, Meta). These dependencies are standard for AI security applications and do not present unusual risks.
[DATA_EXFILTRATION]: Analysis of the source code confirms that input data is processed locally to generate security verdicts. There is no evidence of unauthorized network transmission, data exfiltration, or hardcoded sensitive information.
[COMMAND_EXECUTION]: The skill utilizes pytesseract to extract text from images, which involves executing the Tesseract OCR engine. This is a documented requirement for its core function of identifying text-based attacks rendered into pixels.
[SAFE]: The skill includes an explicit normalization logic to mitigate common attack vectors:
Ingestion points: Processes HTML, PDF, images, and raw text via CLI arguments in agent.py.
Boundary markers: Serves as a pre-ingestion validation gate.
Capability inventory: Limited to local file reading and JSON output; no hazardous shell or network capabilities are exposed to the input data.
Sanitization: Implements character normalization and de-obfuscation (Base64/ROT13) to expose hidden payloads.

detecting-indirect-prompt-injection