Skill UI

Enables agents to perform authorized security assessments, blue team exercises, and penetration testing on AI products, agents, and infrastructure. Uses first-principles reasoning to identify vulnerabilities like prompt injection and tool abuse, generating evidence-based security reports without relying solely on payload libraries.

This skill involves testing AI agents for direct prompt injection vulnerabilities where user messages attempt to override system instructions. It focuses on role confusion and instruction hijacking without external content, ensuring agents maintain security boundaries against direct manipulation attempts.

This skill enables the creation of robust evaluation frameworks for AI agent systems. It is used for systematically testing agent performance, validating context engineering choices, and measuring improvements over time. The approach accounts for non-determinism and multiple valid paths, utilizing multi-dimensional rubrics, LLM-as-judge methods, and human evaluation to ensure quality gates and catch regressions before deployment.

Conduct authorized security assessments for LLM applications and AI agents. Covers prompt injection, tool abuse, RAG exposure, memory poisoning, and system prompt extraction following OWASP LLM/ASI Top 10 guidelines. Includes red-teaming, vulnerability mapping, and compliance engineering for AI systems to ensure safety and robustness.