Anthropic’s Mythos model spots software flaw, draws China-based attack attempts

Anthropic’s Mythos model flagged a newly disclosed software vulnerability and produced a working exploit, prompting immediate attention from security teams. The Register notes that exploitation attempts were traced to a China-hosted IP address, as reported by researcher at VulnCheck link.

Mythos’s strength lies in its mathematical reasoning, which allowed it to identify the flaw and generate exploit code faster than conventional manual analysis. The incident underscores how AI-driven tools can augment vulnerability research, delivering rapid proof-of-concepts that help vendors patch issues before widespread abuse.

The China-based source of the traffic illustrates that AI-generated exploits can quickly become a target for nation-state or proxy threat groups, raising questions about responsible disclosure and the need for robust monitoring when deploying such models.

The episode provides a concrete example of AI integration into corporate security strategies, showing both the efficiency gains for defenders and the heightened exposure to adversarial use.