Red-teaming и аудит безопасности LLM

★ 7.3 · security

Безопасность AI-систем (наступательный аспект) is a Claude Code skill that provides a structured offensive checklist for red-teaming AI and LLM systems: prompt injection, jailbreaking, model extraction, training data poisoning, adversarial inputs, memory poisoning, tool chaining escalation, tokenization exploits, and more. Reach for it when you need to red-team an LLM before shipping it to production, when a client asks how easily their LLM-powered product's content filters can be bypassed, when you're investigating whether data can leak through a RAG pipeline or plugins, or when you need to document the full attack surface of an AI agent with tool access. The checklist — sourced from SnailSploit/offensive-checklist — covers system reconnaissance (model, APIs, data sources, plugins), vulnerability assessment using garak and LLMFuzzer, chained exploit scenarios, and post-exploitation paths such as model theft and lateral movement. It also addresses multi-modal models, multi-agent frameworks (AutoGPT, CrewAI, LangChain Memory), virtualization attacks, and layered policy conflicts. The skill's direction is offensive testing, not defensive hardening or monitoring — use a separate skill for the latter. Use only legally, on systems you own or are explicitly authorized to test.