Google's Gemini 4 Carbon AI Hunts and Patches Security Flaws

Google just built an AI model whose entire job is to break into software, find the cracks, and then fix them before anyone else can exploit them — and for now, it's only trusting a few hundred security teams with it.

The model is called Gemini 4 Carbon, and it landed quietly last week as Google's newest "frontier" model, the industry's term for a lab's most capable, most expensive system at any given moment. Unlike a general chatbot, Carbon is tuned specifically for two things: writing secure code and hunting down vulnerabilities, the technical term for a flaw in software that a hacker can exploit to get in where they shouldn't.

What Carbon Actually Does

According to Google's own benchmark results, Carbon can identify a meaningful majority of planted security flaws in test code, successfully exploit a smaller but still significant share of them to prove they're real, and then write a working patch for most of what it finds. Google says that's more than double the success rate of Gemini 4 Argon, its predecessor, which only launched about two weeks earlier and which Code24 covered when it went live for developers.

That jump matters because the same skill set that makes an AI good at finding bugs for a defender also makes it good at finding bugs for an attacker. It's the classic dual-use problem in security research: a tool that patches a zero-day — a vulnerability nobody has seen or fixed yet — is the same tool that could, in the wrong hands, find one first and weaponize it.

Why the Rollout Is So Slow and Deliberate

Google clearly built its launch plan around that risk. Instead of opening Carbon to every developer with an API key, it's going first to a curated group of cyber defenders through something it calls the Fairwind Program, reportedly already running with several hundred partners: national cyber-response agencies, operators of critical infrastructure like power grids and payment systems, and major technology platforms. Everyone else has to wait.

Google also says Carbon is part of a voluntary pre-release safety testing process run with the US government, and that benchmark numbers so far come from tests run without the usual safety guardrails switched on — meaning real-world performance, once restrictions are applied, could look different.

• Carbon's benchmark scores span three tasks: finding vulnerabilities, exploiting them to confirm they're real, and generating a working patch.
• It's rolling out first through the Fairwind Program, not to the general public or ordinary developers.
• Introductory pricing reportedly matches Argon's: $2 per million input tokens and $10 per million output tokens — tokens being the small chunks of text an AI model processes and bills by.
• Wider access for businesses and individual developers hasn't been scheduled yet.

A model that can be trusted to probe live systems for exploitable flaws isn't a party trick — it's infrastructure. Google handing it first to governments and critical-infrastructure operators, rather than the open market, says as much about how seriously it's treating the risk as any benchmark score does.

What It Means for India

India doesn't have a seat at the table yet in any confirmed sense, but it has plenty at stake. The country's banks, UPI payment rails, and the thousands of IT services firms that write and maintain code for clients worldwide are exactly the kind of targets an AI-assisted attacker — or defender — would be most useful against. CERT-In, India's national cyber emergency response team, has spent the last few years pushing companies toward faster vulnerability disclosure and patching; a tool that can find and fix flaws at machine speed is directly relevant to that mission, assuming Indian agencies or firms eventually get access through programs like Fairwind.

There's also a data angle under the DPDP Act, India's data protection law: any AI system scanning production code and infrastructure for weaknesses is, by definition, looking at systems that may process personal data. Indian companies adopting tools like Carbon down the line will need to think about who gets to see what Carbon finds, and how fast — an unpatched vulnerability report sitting in an inbox is itself a liability.

The Bigger Picture

The real story here isn't that Google built a smarter coding model. It's that the company is treating offensive-capable AI as something closer to a controlled substance than a product — gated, logged, handed out to people with a reason to need it. Whether that caution holds once competitors inevitably ship their own version of the same idea is the question worth watching, because the moment one lab opens the gates, the pressure on everyone else to do the same follows fast.