antgroup/adversarial-ai-coding-plugin

3 stars · Last commit 2026-04-27

Adversarial AI Coding: Self-Play Sparring, Offense Meets Defense — Making AI Write Secure Code

README preview

<p align="center">
  <h1 align="center">Adversarial AI Coding Plugin</h1>
  <p align="center"><b>Self-Play Sparring, Offense Meets Defense — Making AI Write Secure Code</b></p>
</p>

<p align="center">
  <a href="#benchmark-results">Benchmark</a> |
  <a href="#quick-start">Quick Start</a> |
  <a href="#roadmap">Roadmap</a> 
</p>

---

## The Problem: The Security Paradox of SOTA LLMs

LLMs are generating production code at an unprecedented pace, but that code harbors significant security risks. Published benchmarks show that even state-of-the-art commercial models **produce vulnerable code up to 48.2% of the time, with top-tier models ranging from 37% to 95.6%** (Source: AutoBaxBench, Dec 2025).

Paradoxically, these same SOTA LLMs excel at vulnerability discovery — Claude Opus 4.6 has uncovered hundreds of security vulnerabilities in open-source projects and even independently developed a full exploit chain for a FreeBSD kernel remote code execution vulnerability.

**Root cause:** When generating code, the LLM's optimization target is *functional correctness*, not *security*. Security is merely an implicit constraint that is easily diluted under pressure to produce working code.

View full repository on GitHub →