Why would Google build its strongest AI and then not give it to everyone?

Google has a new AI that it says is great at spotting hidden weak spots in software. Only a selected group of defenders can use it so far, and the reason says a lot about how the internet is defended.

This article contains affiliate links. If you buy through them, we may earn a small commission at no extra cost to you.

On 28 October 2026, Google introduced a new artificial intelligence called Gemini 4 Carbon. It is not a robot with metal arms. It is a computer program, and in its own announcement Google calls it its strongest model, built above all for cybersecurity and for long jobs that take a long time to finish. Google says Carbon can keep working for days on a hard engineering task, such as tracking down why a huge program breaks and then fixing it. That is a bit like a mechanic who does not stop until the strange noise in your car is gone.

Many new AI tools arrive with a big public launch. This one did not. Google is giving it first to a selected group of people whose job is to protect computers, and it has not said when everyone else can try it. Why does that matter to you? Because the people it is meant for protect things you use every day: Business Today reported that Google's program for these defenders covers areas like healthcare, telecoms, energy and banking networks.

Security experts call a weak spot a vulnerability. The US standards agency NIST describes it as a weakness in a system that could be exploited, which means someone could use it to do something they should not be able to do. Think of a house where one window latch does not quite close. Nothing bad happens until someone notices it.

Finding a vulnerability is a bit like solving a puzzle. Big programs can hold millions of lines of instructions, and one wrong step can open a gap. Forget the movie hacker typing fast in a dark room. Real bug hunting is often slow, careful reading of code, line after line, looking for the one spot where things go wrong.

Here is the tricky part. The same knowledge helps both sides. An explainer from the Cloud Security Alliance puts it simply: an attacker has to build a working way to use a weak spot, and defenders also have to know about the weak spot in order to fix it. It is like a spy and a guard who both study the same map of a building. One wants to sneak in, the other wants to lock the door. Britain's National Cyber Security Centre says the same about AI: it can help attackers research weak spots and build attacks, and it will also help the people who build and run systems keep them safe.

Google reports three test scores for the new model. Remember, these are Google's own numbers from tests Google ran, and we could not find any outside group that has checked them yet.

On a test called SEC-bench Pro, which checks how well an AI finds weak spots, Google reports a score of 88.2 percent. That is like a spelling test where you get only about 12 words wrong out of 100. On a test called CWE-bench, which checks how well an AI repairs broken code, Google reports 71.0 percent, about 7 fixes out of every 10 tries. On a third test, about actually using weak spots, Google reports 54.8 percent, a little better than calling a coin toss correctly.

The tech news site WinBuzzer adds a detail worth knowing. It reported that Google ran these outside security tests without its safety guardrails switched on, so the version that partners actually use might score differently.

Google describes three jobs for defenders: finding real weak spots, checking which ones an attacker could really use so the worst get fixed first, and writing repairs that remove the problem at its root. Why does this matter to you? Many app updates fix problems like these, and faster fixes mean fewer days when the window latch is still open.

So why keep a tool meant for protection away from most people? The answer seems to be time.

Google's own threat researchers at Mandiant have measured how fast attackers start using a weak spot after it becomes known. In their analysis, the average wait was 63 days in 2018 and 2019, about nine weeks, roughly the time from New Year to early March. By 2021 and 2022 it had dropped to 32 days, about one month. In 2023 it was just 5 days, a single school week. That is more than 12 times faster than a few years earlier.

The UK's National Cyber Security Centre writes in its assessment of AI and cyber threats that the time between a weak spot being made public and someone using it has shrunk to days, and that AI will almost certainly shrink it further.

Now look at how Google explains its plan. On the page for its Fairwind Program, Google says powerful AI models can help defenders find fixes quickly, but "in the wrong hands, the same capabilities can become an equally powerful threat." It says early access goes to the most important defenders so they can strengthen their defenses before new threats appear. In other words, Google frames the holdback as a head start. If the race between finding a weak spot and using it is measured in days, the side that gets the tool first might be able to patch the hole before anyone else even sees it.

There is already one small example of what a head start can look like. In 2024 a Google AI project called Big Sleep found a weak spot that attackers could have used in SQLite, a widely used free database program. The fix was made the same day, before the bug ever reached an official release, so users were not affected, Dark Reading reported.

Carbon starts with a selected group of cyber defenders. After that, Google says, access will slowly widen to more trusted partners through Fairwind. Google says more than 650 partners already take part. If you met one of them every day, it would take you almost two years.

Fairwind puts governments and national cyber agencies first, along with companies that run critical services, such as the ones that keep the electricity flowing, and big technology platforms. The rules are strict. Members go through background checks, must log in with phishing resistant multifactor security, may only give access to their own security teams, and may not share or sell it. Google's page says creating malware is not allowed.

Reuters reported that Google is limiting access while it keeps testing its safeguards, and Google says it is also taking part in the US government's voluntary testing of new AI models before release. Google says it wants to offer Carbon to developers, businesses and everyday users "as soon as possible", but it has not given a date.

Google is not the only lab that holds back cyber capable AI. In April 2026, OpenAI made a security model called GPT-5.4-Cyber available at first only to vetted security companies, organizations and researchers, through a program called Trusted Access for Cyber. According to eWeek, OpenAI argued that cyber risk "cannot be judged by model capability alone" and that it matters who is using a system and how much they can be trusted.

Gatekeeping has a cost too. Axios quoted an OpenAI researcher, Fouad Matin, saying: "No one should be in the business of picking winners and losers when it comes to cybersecurity." Small companies and independent researchers around the world defend systems too, and a narrow rollout could leave many of them waiting. RuntimeWire noted that for now most developers simply cannot test how the model performs in real work.

So there are two worries pulling in opposite directions. Hand the tool to everyone, and attackers could get it on day one. Hand it to only a few, and many defenders are left out.

In Google's framing, a head start for trusted defenders comes before a fast public launch. Whether that choice pays off depends on things nobody can see yet: how well Carbon works outside Google's own tests, how quickly the circle of users grows, and whether attackers build similar tools of their own.

If an AI can find weak spots faster than people can fix them, who should get it first, and who should decide?