Anthropic's Mythos can convert software vulnerabilities into exploits quickly

Anthropic's Mythos Preview can transform newly disclosed software vulnerabilities into working exploits within hours, as detailed in new research shared with Axios. The findings indicate that advanced AI models can effectively weaponize known flaws, potentially reducing the time between vulnerability disclosure and patching. Anthropic's red team tested Mythos against vulnerabilities in Mozilla Firefox and the Microsoft Windows kernel disclosed in January and February. Within 31 minutes, Mythos produced its first proof-of-concept exploit for a Windows kernel vulnerability. Out of 21 kernel bugs tested, it caused a 'blue screen of death' in 18 cases and created eight distinct exploits, with the longest taking approximately 5.7 hours to develop. For Firefox, Mythos successfully built eight working code-execution exploits from 18 security patches. Most cyberattacks exploit known vulnerabilities that companies have not yet patched. Patching systems often requires testing to prevent crashes and may involve downtime. Other open-source models are reportedly finding bugs at levels comparable to Mythos and OpenAI's GPT-5.5-Cyber. Anthropic estimates the cost to generate Windows privilege-escalation exploits at about $15,700 in API credits, or roughly $2,000 per exploit. The Trump administration is implementing a new AI security executive order to assess national security risks associated with advanced AI models.