本文へスキップ
Kitbay
スキル

llm-security

MIT2026年11月14日 更新44,947GitHubスター

A Chinese-language skill from the reverse-skill collection for authorised security assessments of LLM applications and AI agents, mapped to OWASP's lists of risks for LLM and agentic applications.

スキルの内容

A Chinese-language skill from the reverse-skill collection for authorised security assessments of LLM applications and AI agents, mapped to OWASP's lists of risks for LLM and agentic applications. The agent maps the attack surface: inputs, registered tools, data flows through retrieval and tool calls, system prompt leaks and human approval points. It then tests prompt injection from direct overrides through role-play, encoding tricks and multi-turn extraction to indirect injection via web pages, PDFs or email, tool abuse and approval bypass, poisoning of retrieval and long-term memory, unsafe use of model output downstream, and system prompt extraction checked with canary tokens. It names garak, PyRIT, promptfoo and similar tools, and asks for reproducible evidence.

できること

  • Map the attack surface of an AI application
  • Test direct and indirect prompt injection
  • Check tool abuse and approval bypass
  • Test retrieval and memory poisoning