The Agent Skills Directory

[PROMPT_INJECTION]: The file contains a collection of prompt injection and jailbreak strings (e.g., 'Ignore previous instructions', 'Act as DAN') intended for use in security testing and vulnerability assessment of AI models.
[DYNAMIC_EXECUTION]: The skill highlights security risks associated with untrusted deserialization using pickle.loads() and the execution of system commands via os.system() or subprocess within machine learning pipelines.
[REMOTE_CODE_EXECUTION]: The documentation provides examples of how malicious code can be executed during model loading, such as using shell commands to load pickled files, and suggests tools like safety for scanning directories.
[DATA_EXFILTRATION]: The skill illustrates potential exfiltration vectors, including URLs used by attackers to send data from compromised RAG pipelines, serving as a guide for defensive monitoring.

ai-security