Written By | Torch Media Africa
Gemini 4 Carbon tops the AI leaderboards and can find, test and patch software flaws on its own for days at a time. So why won't Google let you near it?
Search
TECH & SCIENCE DAILY
Google has built an artificial intelligence that can spend days hunting through computer code for hidden weaknesses, work out whether a criminal could exploit them, and then write the repair itself. And right now, almost nobody is allowed to use it.
The machine is called Gemini 4 Carbon, and it was unveiled this week by Google DeepMind, just weeks after the company launched its sibling model, Gemini 4 Argon. According to Reuters, Google calls Carbon its latest frontier AI model and says it performs at state-of-the-art levels on cybersecurity tasks, including finding vulnerabilities, exploiting them and patching software. It is also, on at least one widely watched yardstick, the smartest AI on the planet.
But there is a twist that makes this launch very different from the usual parade of shiny chatbots. Carbon is not going to your phone, your laptop or your office tools. Not yet. Initial access is limited to a select group of cybersecurity defenders, and Google says it will widen the circle slowly, testing safeguards as it goes. In a world where the same skills that protect a hospital network could help break into one, the guest list is the story.
What Gemini 4 Carbon actually does
Most AI models are generalists. Carbon is something closer to a specialist with a very long attention span. Coverage of the launch describes it as tuned for long, multi-day reasoning loops: it scans code for vulnerabilities, tests whether each one can really be exploited, and then drafts a fix. Google says it is built for complex, long-running engineering and coding tasks, and describes it as tailored to defensive security and autonomous software remediation.
Think of it as a tireless security guard who never sleeps, never gets bored of reading millions of lines of code, and can pick up a case on Monday and still be working it on Thursday. Human security teams are famously outnumbered and overworked. Software flaws pile up faster than people can patch them. An AI that can shoulder the grinding middle of that job, finding the bug, proving it matters and proposing the repair, is exactly what defenders have been dreaming about.
The numbers Google is publishing are eye-catching. On SEC-bench Pro, a test of whether AI can spot real software flaws, Carbon scored 88.2 percent. On CWE-bench, which measures whether a system can produce working fixes for categories of known weakness, it hit 71.0 percent. On ExploitGym, which tests whether discovered flaws can actually be exploited, it managed 54.8 percent. Google also says Carbon produces viable fixes at more than twice the rate of Argon, and that its Chrome team measured patch quality more than ten times better than an unnamed rival frontier model.
Those are vendor-reported figures, and they should be read with the usual caution. Independent scrutiny is what turns a press release into a fact, and so far the most important outside verdict concerns Carbon's general brainpower rather than its hacking-hunting skills.
The leaderboard shocker
Carbon has landed at number one on the Artificial Analysis Intelligence Index v5.0, an independent ranking of AI models, with a score of 59. That edges past Claude Opus 5.5, which sits at 58, and OpenAI's GPT-6 Astra. In Google's own tally, Carbon led 15 of 17 standard benchmarks.
It was not a clean sweep. Carbon trailed rivals in software engineering, measured by FrontierSWE 2, and in computer use, measured by OSWorld-2.1. For a model sold on its coding and engineering endurance, falling behind on a software engineering test will raise eyebrows among developers, and it is a reminder that benchmark crowns change hands quickly in this race.
There is also a blemish in the honesty department. Artificial Analysis ranked Carbon first on its AA-Omniscience test with a score of 54, but the model's hallucination rate, meaning how often it confidently states something unsupported, rose to 11 percent from Argon's 4 percent. In fairness, reporting on the launch notes that this is still far below the roughly 50 percent rates recorded for Claude Opus 5.5 and GPT-6 Astra on the same test. But for a tool whose job is to tell defenders what is wrong with their code, a made-up answer is not a harmless quirk. A false alarm wastes a security team's day. A missed or invented fix could do real damage.
Why the velvet rope?
Here is the part that has the tech world talking. A model that can find a flaw, prove it can be exploited and write the repair is, by definition, a model that understands how to break things. Security people call this dual use. The knowledge that lets a defender patch a hole is the same knowledge that lets an attacker walk through it.
Google's answer is a staged rollout. Carbon is going first to a limited group of cyber defenders, starting this week. From there Google plans to expand access through the Fairwind Program, its scheme for putting the most capable models into the hands of the people who guard critical systems. Government bodies, critical infrastructure operators and enterprise defense teams are in line. Google says it will gather feedback and refine its safety controls before any wider release on Vertex AI, the Gemini API and the Gemini app. No general availability date has been announced.
The company is also taking part in a voluntary pre-release testing process run with the United States government, a sign that the people who build these systems and the people who regulate them are starting to treat the most capable cyber models a little like sensitive technology rather than ordinary software.
Inside the Fairwind Program
Details published around the earlier Argon launch show how tightly the door is guarded. Google DeepMind describes Fairwind as its program for giving its most capable models early to those who defend critical infrastructure, and says it already works with more than 650 partners.
Access follows a priority order: governments and national cyber authorities first, then operators of critical infrastructure such as healthcare, telecoms, energy and finance, then large technology platforms. Google checks applicants' security history and conduct. Every user needs an individual account protected by phishing-resistant multi-factor authentication. The model may only be used by internal cybersecurity, incident response or penetration-testing teams, organizations must track who is using it and how, and access cannot be shared, resold or redistributed.
Dual-use work that defenders genuinely need, such as authorized attack simulations and malware analysis, is allowed. Creating malware is explicitly off the table.
One detail raised eyebrows: for admitted partners and Google's own internal teams, the earlier Argon model is described as being available without its usual cyber guardrails. In other words, the safety brake moves from the model to the organization using it. Trust, in this system, is a membership requirement.
The safety scorecard
Google is leaning hard on its safety numbers. According to its launch material, Carbon recorded zero successful attacks on Gray Swan's indirect prompt injection benchmark, an improvement on Argon's 0.7 percent. Prompt injection is the trick of hiding malicious instructions inside text an AI is reading, and it is one of the biggest unsolved weaknesses of AI agents that browse, read email or process documents.
Google also reports that Carbon takes 26 percent fewer actions outside user-defined limits and abandons 29 percent fewer multi-step tasks than Argon. On Terminal-Bench 4.0, a test that watches for models cheating their way through tasks, environment exploits and rule-breaking shortcuts fell to 0.0 percent. Internal testing showed unsupported claims roughly halved against Argon.
These are internal results. One industry digest covering the launch pointed out that the claims still need outside scrutiny, and that the access decisions may reveal more about how Google views the model's power than any benchmark chart.
The price tag
For the few who can get in, the pricing is aggressive. Via the Gemini API, introductory rates are $2 per million input tokens and $10 per million output tokens, the same as Argon's introductory pricing. Those rates are due to double, to $4 and $20, later on. Cached input gets a 95 percent discount, which matters enormously for security work that repeatedly reads the same huge codebases.
Cheap, tireless code review is a serious proposition. A task that would take a human team weeks of billable hours can be left running for days at a cost most companies would barely notice. That is exactly why Google's rivals are watching.
A race with OpenAI and Anthropic
Reuters framed the launch as stepping up competition with OpenAI and Anthropic as the race to build more capable coding and cybersecurity systems intensifies. The crowded leaderboard shows how tight it has become. A single point separates Carbon from Claude Opus 5.5 at the top of the Artificial Analysis index, and each lab is racing to prove its model is not just clever but dependable enough to be left alone with a company's most sensitive code.
Cybersecurity has become the prestige battleground for a simple reason. It is a field where an AI's abilities can be measured against hard, checkable results, a vulnerability is either real or it is not, a patch either works or it does not, and where the commercial and national-security stakes are enormous. A model that can reliably find and fix flaws becomes a must-have for governments and corporations alike.
The expert view: promise and peril
Neither Google's announcement nor the Reuters report carried quotes from outside experts, so the clearest outside reading comes from the analysts who have dissected the launch. Their consensus, put plainly: the technology could be a genuine gift to overstretched defenders, but the same capability is a double-edged sword. Cybersecurity models can help find and fix flaws, one widely read AI digest observed, but the same capabilities could help someone exploit them.
That tension explains the cautious rollout. Security researchers have long argued that defenders are structurally disadvantaged, because an attacker only has to find one way in while a defender has to close every door. A tireless AI that levels that playing field is welcome news. But if the equivalent tool ever leaked, was stolen or was replicated by a less careful developer, the advantage could flip overnight. Reports this week of advanced AI chips being smuggled through Malaysia, more than $12 million worth hidden in 30 servers at Kuala Lumpur airport, show that the hardware race behind all of this is already a cat-and-mouse game.
What this means for you
You will not be typing questions into Gemini 4 Carbon this weekend. But you may well feel its effects. If it performs as advertised, the apps, banking systems and hospital networks you rely on could get patched faster, with fewer long-lived holes sitting open for criminals to find. The software updates that pop up on your phone may increasingly be written, at least in part, by machines that found the bug first.
The flip side is that attackers are expected to get smarter tools too. Security experts have been telling ordinary users the same few things for years, and they matter more, not less, in an AI era. Turn on multi-factor authentication, preferably with a passkey or security key. Install updates promptly. Use a password manager. Be suspicious of unexpected messages, because AI is making scams more convincing.
For workers, the message is mixed. Entry-level security analysts who spend their days triaging alerts and reading code may find parts of their job automated. Skilled professionals who can direct, check and trust these systems are likely to be in more demand, not less. Google's own hallucination numbers are a good reminder that someone has to check the machine's work.
The bottom line
Gemini 4 Carbon is a two-sided headline. On one side sits a model that tops an independent intelligence ranking and claims to fix software flaws at more than twice the rate of the model released weeks earlier. On the other sits a company deliberately keeping it away from the public because of what it could do in the wrong hands.
That caution may be the most telling thing about the launch. The leading AI labs are now building systems powerful enough that they decide who gets access before deciding what the product is. If Google's safeguards hold, Carbon could be remembered as the moment AI finally helped the defenders catch up. If they do not, the velvet rope will have been a warning that nobody heeded.
For now, the doors stay closed, the testing continues, and the world waits to see whether the guard that never sleeps can be trusted with the keys.
