About shimo4228/llm-agent-security-principles
An Agent Skill holding the structural security playbook for autonomous LLM agents, for developers who design an agent, add a capability to one, or review one for prompt-injection resistance. Web-app security controls (the OWASP web Top 10, auth, input validation) do not cover the threats that matter most here: prompt injection, injected text that the agent's own memory carries into every later prompt, and the damage that grows with each outside system one agent can reach. The core insight: you cannot reliably restrict an LLM's behavior at runtime. You must shape the system so dangerous behaviors are structurally impossible.