ai-security

Prompt injection is an attack technique where malicious inputs manipulate LLM behavior to bypass safety guardrails or leak sensitive company information. This Skill simulates prompt injection attacks to identify such vulnerabilities in AI systems before deployment.

To assess LLM vulnerabilities, this Skill conducts offensive security testing using the OWASP Top 10 for LLM Applications and MITRE ATLAS methodologies. It identifies weaknesses like insecure output handling, excessive agency, and data leakage in AI systems.

Yes, this Skill is designed for red teaming AI and ML systems by simulating adversarial attacks. It tests for model extraction, training data poisoning, and denial-of-service vulnerabilities to evaluate system resilience against advanced threats.

The best way to test for model extraction and data poisoning is through targeted attack simulation. This Skill employs adversarial techniques to attempt data leakage and model parameter extraction, evaluating your AI system's defenses against these specific threats.

Yes, this Skill supports testing jailbreaking and excessive agency in LLMs. It evaluates whether LLM-assisted applications can be manipulated into executing unauthorized commands or bypassing intended operational constraints during red team simulations.