Researchers have successfully replicated dangerous AI vulnerability findings, originally identified by Anthropic's advanced 'Mythos' model, using publicly accessible AI tools. This suggests that the ability to discover AI-driven security exploits is becoming inexpensive and widely available.
The study, conducted by Vidoc Security, used models like GPT-5.4 and Claude Opus 4.6 within an open-source coding agent. They reproduced several bug cases that Anthropic had highlighted as critical, including vulnerabilities in file-sharing protocols, operating system networking stacks, video processing software, and cryptographic libraries. The cost for each scan remained under $30.
While the researchers' AI found flaws in code, they did not build complete attack blueprints, indicating a gap between vulnerability discovery and exploitation. However, the core finding emphasizes that the economics of vulnerability discovery are shifting, moving from exclusive model access to broader validation capabilities.