Measuring LLMs' impact on N-day exploits
Given recent patches, Mythos Preview autonomously built 8 working exploits from 18 Firefox patches and 8 SYSTEM-level chains from 21 Windows kernel patches.
Studies the patch gap, where models diff a published fix to reverse-engineer the bug. Public Claude models with safeguards off can also build exploits, though fewer than Mythos Preview. Anthropic urges defenders to deploy patches faster because exploit development, once the scarce-expertise step, has largely stopped being a bottleneck.
- Date
- Monday, 8 June 2026
- Lab
- Anthropic
- Kind
- paper
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Firefox patches turned into code-execution exploits | 8 of 18 Claude Mythos Preview, autonomous | company |
| Windows kernel patches turned into full exploit chains | 8 of 21 low-privilege user to SYSTEM; no source code available | company |
Company measurement with safeguards off; real campaigns also need target discovery and delivery.
Sources
This record was checked against its sources on 6 October 2026. How we check