Google's new model tops the index at finding flaws in code, so Google is locking it to defenders first

Two weeks after its last flagship, Google DeepMind has shipped another one, and this one's speciality is breaking into things. Gemini 4 Carbon, launched on Wednesday two weeks after the model called Argon, has gone straight to first place on the Artificial Analysis Intelligence Index v5.0 with a score of 59. The benchmarks underneath that headline are unusually specific: 88.2 percent on SEC-bench Pro vulnerability tests, 54.8 percent on ExploitGym and 71.0 percent on CWE-bench. Google says its Chrome team found the model patches code more than ten times better than a rival frontier model it tested, and the launch reporting puts Carbon at more than twice as good as its predecessor generation at generating secure code.

The interesting decision is not the model. It is the door. A system this capable at finding and exploiting software flaws is dual use in the purest sense: the same skill that lets a defender find a hole before it is used lets an attacker find it first. Google's answer is the Fairwind Program, a staged rollout that gives first access only to vetted cyber defenders, at introductory pricing of 2 dollars per million input tokens, with the wider release following later. DeepMind's Koray Kavukcuoglu framed the launch around exactly that tension, a model built to strengthen the people defending software, released in an order designed to keep it that way.

Why Google is moving this fast sits in the calendar. Two flagship class releases in a fortnight is not a research rhythm, it is a competitive one, aimed at a market where the lead on the public indexes changes hands monthly and where enterprise buyers in security are among the few customers paying serious money for frontier capability today. Taking first place on a respected independent index with a security specialist model does two jobs at once: it wins the general race for a news cycle, and it plants Google's flag in the one AI category where the buyer, the security team, has both the budget and the urgency.

The staged rollout will be tested by the obvious question: how long the moat holds. Vetting defenders works while the capability is scarce, and capability in this industry has a half life. Rival models will post similar exploit scores, open models will follow them, and the window in which Fairwind's careful sequencing is the binding constraint on who can use this power is measured in months, not years. Google knows that, which is why the program reads less like a permanent gate and more like a head start, bought deliberately for the defensive side.

Our call is that the staged release becomes the industry template for dual use models, because the alternative positions are both worse. Releasing a flaw finding model of this strength to everyone on day one invites the headline nobody in this industry can afford, and never releasing it hands the capability to whoever builds it next with fewer scruples. Watch whether rival labs copy the vetted defender structure within their next two launches, and watch Fairwind's vetting list for the first argument about who counts as a defender. That argument, not the benchmark score, is where the real governance of these models is going to be decided.