Call
All work

// open source

mii-ai-security

Think like an attacker. Ship like a defender.

Stack
Python · machine-readable skill index
Distribution
MIT · npm: miii-security
Coverage
58 skills · 12 domains
GitHub

In plain words

AI chatbots can be tricked into leaking data or misbehaving. This is a free, practical playbook that shows engineering teams how to prevent it.

Think
Teams ship AI faster than their security habits mature, and most advice is too vague to act on.
Design
58 short guides, each with how an attacker thinks, what to fix, and quick wins, written for the tools teams already use.
Build
Mapped to OWASP LLM Top 10 and MITRE, with a machine-readable index, validation, and red-team test scripts.
// 00 — The full story, for engineers

§ 01

The problem

Teams ship production LLM systems faster than the security practices around them mature. There is plenty of high-level advice about prompt injection and data leakage, and very little that an engineer can actually pick up and apply during a review.

§ 02

The decisions that mattered

  • Structure security as skills, not prose. 58 SKILL.md guides, each with an attacker mental model, a control table with severity ratings, and quick wins you can ship today.
  • Make it framework-native. Controls are written for LangChain, LlamaIndex, Semantic Kernel, AutoGen, CrewAI and others, because generic advice dies on contact with a real codebase.
  • Map everything to OWASP LLM Top 10 and MITRE, so the work plugs into how security teams already think.
  • Ship a machine-readable index, validation pipeline, adversarial fixtures, and red-team scripts, so the framework is testable, not just readable.

§ 03

Why it matters to me

I do AI security as part of my day job at a cybersecurity company. This is me turning that into something portable: the review I wish every team running LLMs in production had open in a tab.

Want this kind of thinking on your product?

Book a 20-min call