Google's Gemini 4 Carbon Leads Vulnerability Benchmarks

Google has unveiled Gemini 4 Carbon, its new flagship model. The company claims state-of-the-art results in cybersecurity and in long-horizon tasks where the model works on its own for extended periods.

On SEC-bench Pro, which tests finding vulnerabilities in code, Gemini 4 Carbon scored 88.2%. On ExploitGym, which tests exploiting them, it reached 54.8%. The source headline also claims a lead in patching vulnerabilities, but the available excerpt gives no figure for that test.

This matters beyond programmers. A model that finds and fixes holes helps defenders. A model that exploits them successfully about half the time also helps attackers, and the gap between a bug being found and mass attacks could shrink.

There is a catch: these are Google’s own claims and benchmark results. Tests show potential, not what happens on real servers. Such numbers deserve caution until independent teams verify them.

Source: The Hacker News