Project Glasswing: What We Learned Running a Frontier Model on Our Own Codebase

Published

2026-09-01 14:53

Collected

2026-09-02 00:15

---
id: 7286
title: "Project Glasswing: What We Learned Running a Frontier Model on Our Own Codebase"
source_domain: hackerone.com
source_url: "https://www.hackerone.com/blog/project-glasswing-frontier-model-h1"
published_at: "2026-09-01T14:53:58Z"
tags: ["ai-security", "rce", "vulnerability-research", "hackerone", "frontier-models", "compositional-risk"]
phying_url: "https://phying.com/item/7286"
lang: en
---

# Project Glasswing: What We Learned Running a Frontier Model on Our Own Codebase

## Summary

HackerOne ran Anthropic's Mythos 5 model against its own production code; it surfaced a critical RCE spanning three safe-in-isolation commits, illustrating compositional risk that diff review misses.

## Intelligence Metadata

- **Sources:** hackerone.com
- **Published:** 2026-09-01T14:53:58Z
- **Original:** https://www.hackerone.com/blog/project-glasswing-frontier-model-h1
- **Phying News:** https://phying.com/item/7286