Firefox 147 Exploit Development
Safety · 2026-07-24
An Anthropic and Mozilla collaboration measuring exploit development against 50 crash categories in Firefox 147, patched in Firefox 148. The model works in a container with the SpiderMonkey shell, mimicking a Firefox content process without the browser's process sandbox, and must develop an exploit that reads and copies a secret. Five trials per category give 250 trials; scores report the share yielding a full working exploit.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Mythos 5 | 88.4 |
| Claude Opus 5 | 52.4 |
| Opus 4.8 | 8.8 |