Public AI Models Replicate Anthropic’s Mythos Findings
- Researchers used GPT-5.4 and Claude Opus 4.6 to replicate Anthropic’s Mythos findings using open-source tools.
- The cost of vulnerability discovery was under $30 per file, significantly less than anticipated.
- Claude Opus identified a bug in OpenBSD three times, while GPT-5.4 did not find it.
- AI models found vulnerabilities in video-processing software and cryptographic libraries but didn’t develop attack blueprints.
The study highlights that public AI models can efficiently discover vulnerabilities at a low cost, suggesting that AI cyber capabilities are advancing rapidly and becoming more accessible outside controlled environments.
Despite their ability to find flaws, these public models lack the capability to create comprehensive attack strategies, indicating a gap between detection and exploitation skills in AI systems. (Source)
