aegis-llm-ai
Installation
SKILL.md
LLM / AI Feature Security Checks (LLM)
Scope & Objectives
- Cover every code path where a large language model (LLM) is invoked from application source: Python (
openai,anthropic,langchain,llama-index), JavaScript/TypeScript (openainpm package, Vercel AI SDK@ai-sdk/*,langchain.js), and raw HTTP integrations hitting provider endpoints. - Cover supporting assets: prompt templates (
.txt,.md,.j2/Jinja, string constants), agent/tool registries and function-calling schemas, RAG pipelines (loaders, extractors, embedders, vector stores, retrievers), conversation-history persistence, and model/provider configuration. - Cover fine-tune/training pipelines ONLY where datasets or training jobs are managed in-repo (poisoning-surface review).
- Out of scope: model internals and weights, provider-side infrastructure security, jailbreak research beyond what the application enables, non-LLM chat systems. Legal determinations about data-processing agreements are out of scope — report findings factually; defer interpretation to counsel.
- Deliverables per finding: call-site location, taint path (source to prompt or completion to sink), trust verdict per interpolated field, executing principal for tools, exploitability reasoning, concrete fix.
- Assume code-read access only. Dynamic payloads in Exploitation & Reproduction are TEST payloads for authorized verification against products you own or are contracted to assess — never against third-party systems.
Objectives, keyed to OWASP Top 10 for LLM Applications category NAMES (numbers deliberately omitted):