Security

Aegis

An AI-security learning platform: prompt injection, tool poisoning and MCP risk, with labs where you attack and then detect.

Status: Working prototype
Role: Designer and developer.
Year: 2026
Stack: Next.js, TypeScript, Vitest, Expo (mobile, in progress)

Aegis lesson listing prompt injection techniques with ATLAS identifiers — Real capture. A lesson on prompt injection techniques.

Aegis labs list including a prompt injection detection lab — Real capture. A lesson on prompt injection techniques.

The problem

Security teams are adopting LLM tools faster than they are learning how those tools fail. I wanted a structured way to learn the attacks and practise detecting them in logs.

Architecture and implementation

12 lessons covering prompt injection, indirect injection, jailbreaks, prompt leakage, system-prompt security, tool poisoning, MCP security, RAG, memory and agent security.
3 labs: break a deliberately vulnerable chatbot, detect prompt injection in chatbot, RAG and tool-call logs, and audit MCP tools for poisoning.
2 real-incident case files, and technique mappings to MITRE ATLAS identifiers.
SIEM-style labs run fully offline with no API key.

Decisions and challenges

Offline first: Labs use bundled datasets so anyone can practise detection without a model account or spend.
Tests as a contract: Lesson content and lab scoring are covered by an automated suite so changes can't silently break a lab.

Verified outcomes

Test suite: 306 of 306 passing across 40 files (run on 25 October 2026).