Google and Anthropic push AI models that find and exploit vulnerabilities

· Cybersecurity · United States

Google unveiled Gemini 4 Carbon, a new frontier model it says leads benchmarks for cybersecurity work, scoring 88.2% on SEC-bench Pro and showing gains in coding, agentic tasks and long-horizon capabilities. Google is rolling it out first to vetted defenders through its Fairwind Program. Anthropic, meanwhile, said an unreleased version of its Claude model found and chained multiple vulnerabilities across Chrome, Safari, Android, Windows and Linux, and that on October 27 its model wrote an exploit for a critical Linux kernel TCP flaw. Both disclosures mark a step change in how quickly AI can turn software weaknesses into working exploits, putting pressure on defenders to shorten patch cycles.

Why it matters：Frontier models are becoming vulnerability-research tools in their own right, shifting the balance between attackers and defenders and raising questions about who gets access to offensive-grade AI.

Key points

  • Gemini 4 Carbon scored 88.2% on SEC-bench Pro, Google says
  • Access is limited first to vetted defenders via the Fairwind Program
  • Anthropic says an unreleased Claude model chained bugs across major browsers and operating systems

Sources (original reporting)

More stories from this day