plimsolllisted

Security review of LLM applications and agents. Use when asked to review, red-team, threat-model or test an LLM feature, chatbot, RAG pipeline, MCP server or autonomous agent for prompt injection (direct or indirect), jailbreaks, system prompt or secret leakage, tool abuse, excessive agency, confused-deputy behavior, authorization bypass, RAG poisoning, memory poisoning, or data exfiltration. Also use when asked to build a red-team test matrix, judge whether a suspicious model response is a real vulnerability, score a finding, or write a regression test for one.
lumiboi/plimsoll · ★ 2 · AI & Automation · score 78

Install: claude install-skill lumiboi/plimsoll

# LLM and agent security review Work through a system the way an attacker reaches it: find the text they control, find the privileges the agent holds, and find out what stands between the two. The one rule everything else follows from: **Never rely on the model to protect a boundary that deterministic software can enforce.** A recommendation that amounts to better prompt wording is not a fix. It may reduce how often the model misbehaves; it does not change what a misbehaving model can reach. ## When this applies Reach for this skill for: agent and LLM security reviews, prompt-security questions, RAG and memory trust analysis, tool and MCP permission audits, red-team planning, triage of a suspicious transcript, and regression tests for a fixed issue. Not for: model evaluation quality, content moderation policy, or general application security unconnected to a model. ## Before touching anything 1. **Confirm authorization.** Ask who owns the target and what is in scope. If the answer is not clear, do the architecture review from source and configuration only, and hold the adversarial tests. 2. **Prefer a sandbox.** Local, staging or a mock-tool harness. If the only instance is production, run the architecture review there and reproduce anything else elsewhere. 3. **Use synthetic secrets.** `plimsoll canary new`. Never a real credential, never real customer data. 4. **Keep the smallest proof.** Once a finding is demonstrated, stop. One mock email in a si