The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents various prompt injection and jailbreak payloads (e.g., 'Ignore previous instructions', 'Act as DAN'). These are presented as educational test cases for security professionals to evaluate third-party LLM applications.- [REMOTE_CODE_EXECUTION]: Illustrates risks associated with unsafe deserialization in machine learning models, specifically demonstrating how the 'pickle' library can be exploited during model loading.- [COMMAND_EXECUTION]: Identifies shell execution patterns like os.system() and subprocess as high-risk indicators to monitor when auditing model source code or supply chain components.- [EXTERNAL_DOWNLOADS]: Discusses security considerations for external model hubs like Hugging Face, including guidance on verifying model integrity and scanning for vulnerabilities.

ai-security