Google says Gemini 4 Carbon is the first frontier model it has tuned explicitly for finding security flaws, judging whether an attacker could reach them, and writing the patch. It arrives with an access lane reserved for defenders.
The headline claims are internal. On Google’s own benchmarks, Carbon generated secure code more than twice as often as Argon, its predecessor. The Chrome team reported it produced high-quality fixes more than ten times better than another lab’s large frontier model.
External results are narrower and more useful. Carbon scored 88.2% on SEC-bench Pro for locating vulnerabilities in real software, 54.8% on ExploitGym for deciding exploitability, and 71.0% on Collinear’s CWE-bench for remediation. Google notes the evaluations ran without guardrails, so those figures describe raw capability rather than what a customer will see.
The lab reports are the interesting part
Internal teams watched Carbon spend days inside large codebases and low-level systems working through a single bug. Koray Kavukcuoglu, Google’s chief AI architect, framed the gains as reduced misalignment and fabrication alongside the secure-code uplift, which is the claim that matters if you plan to let an agent touch production repositories.
Wiz, the cloud security platform Google absorbed, reported Carbon topping its web penetration testing benchmarks. That is AI on both sides of the assessment: the same class of model that finds flaws is now grading the finder.
What a CISO should take from this
The safety stack carries over from Argon: training-time work, live misuse defences, prompt injection resistance and reasoning monitoring. None of it replaces review.
Treat vendor benchmark claims as directional. Pull your own backlog through the model against a fixed corpus, measure false positives before trusting a triage queue, and keep a human gate on patches destined for anything internet-facing.
