1---2name: ai-redteam3description: AI and LLM red-teaming skill for evaluating prompt injection, jailbreak, data exfiltration, tool abuse, agent hijack, RAG poisoning, model denial of service, and unsafe tool-use chains in Gemini, Claude, GPT, and open-weight models. Use to build evaluation harnesses, attack corpora, defensive guardrails, and red-team reports for AI systems you own or are authorized to test.4metadata:5 source: https://skillmd.com/skills/masriyan/ai-redteam6---78# AI Red Team910## Authorization Boundary1112- Only target models, agents, and applications the user owns or has written authorization to test.13- Do not produce working malware, CSAM, weapons synthesis, or content that bypasses model safety for harmful real-world outcomes.14- Treat findings as defensive: every successful attack must ship with a detection and a mitigation.1516## Attack Taxonomy1718- `Direct injection`: instructions in user input override system policy.19- `Indirect injection`: hostile content in retrieved docs, web pages, emails, files, image alt-text, or tool output.20- `Jailbreak`: roleplay, hypothetical, encoding, obfuscation, multi-turn drift, persona pinning.21- `Tool abuse`: forcing browse, shell, code, or MCP tools to act outside scope (SSRF, file read, command exec).22- `Data exfiltration`: leaking system prompt, secrets, embeddings, training data, or per-user memory.23- `Agent hijack`: rewriting plans, looping, escalating permissions, calling unintended sub-agents.24- `RAG poisoning`: index pollution, ranking manipulation, embedding collisions.25- `Resource exhaustion`: token bombs, recursive tool calls, prompt-bomb DoS.2627## Evaluation Workflow28291. Map the AI surface: model, system prompt, tools, retrievers, memory, output sinks, and downstream callers.302. Build a threat model with abuse cases per surface.313. Generate a versioned attack corpus (`attacks.jsonl`) with `{id, technique, payload, expected_block, severity}`.324. Run the harness; record `{pass, fail, partial, refusal_quality, leak_evidence}`.335. Score with attack success rate per technique, mean time to refuse, and blast radius.3435## Defensive Output3637- Input/output classifiers and allow-listed tool schemas.38- Provenance tags on retrieved context; quarantine for untrusted sources.39- Plan-then-execute pattern with human approval on high-impact tools.40- Rate limits, recursion caps, and per-tool budgets.41- Red-team report: top techniques, reproductions, fix owners, regression tests.4243## Deliverables4445- `surface.md`, `threats.md`, `attacks.jsonl`, `harness/`, `results.csv`, `report.md`, `regression-suite/`.