Skip to content

Aegis

Aegis is an autonomous purple-team engine designed to evaluate, exploit, and self-harden AI agents in a closed execution loop.

Code Video Cerebral Valley

Core trick

Parallel adversarial sub-agents mutate OWASP LLM attacks against a target bot, then synthesize runtime prompt patches from breach telemetry.

Tags

securityagents

Description

Aegis is an autonomous purple-team engine designed to evaluate, exploit, and self-harden AI agents in a closed execution loop. Built from scratch with zero framework bloat, Aegis deploys parallel adversarial sub-agents to target a helper bot with mutated context-manipulation and tool-abuse exploits (OWASP LLM Top 10). When a vulnerability is exposed, a defensive orchestration layer dynamically analyzes the breach telemetry, synthesizes targeted runtime prompt-boundary patches, and updates the agent's context in real-time. The entire process streams into a custom interactive 3D telemetry dashboard (dashboard.html) showing the live transition from systemic breach to a hardened 3/3 blocked security posture.
gadaugherty/Aegis★ 10 forksHTMLMITCreated May 23, 2026 · last push May 24, 2026

README excerpt

# Aegis — Autonomous Purple-Team Swarm

**Built at Google I/O 2026 Hackathon · Powered by Gemini 3.5 Flash**

> Every team here shipped an AI agent. Almost none tested it for prompt injection.
> Aegis is a swarm of Gemini 3.5 Flash sub-agents that **breaks an AI agent and then hardens it — autonomously, in a closed loop.**

<div align="center">
  <a href="https://www.loom.com/share/d7bc86fcc85f41e295fca588c5753db0">
    <h3>Aegis Evaluates and Self Hardens Agents — Watch Demo</h3>
  </a>
  <br />
  <a href="https://www.loom.com/share/d7bc86fcc85f41e295fca588c5753db0">
    <img src="https://cdn.loom.com/sessions/thumbnails/d7bc86fcc85f41e295fca588c5753db0-59bb710e4d450eba-full-play.gif" width="700" alt="Aegis Demo Walkthrough" />
  </a>
</div>

---

## The Problem

Red-team tools for LLMs (Garak, PyRIT, promptfoo) *find* holes. They hand you a list of failures and stop. A human still has to read the report, write the fix, deploy it, and re-test by hand.

Aegis closes that loop. It finds the hole, **writes the guardrail, applies it, and re-attacks the hardened agent to prove the fix held** — with no human in the middle.

              ┌────────────────────────────────┐
              │   Target Agent System Prompt   │◀─────────────────┐
              └────────────────────────────────┘                  │
                              │                                   │
                              ▼                                   │
                ┌─────────────────────────

Team