- Google says Gemini 4 Carbon scores 88.2% on finding vulnerabilities, 54.8% on exploiting them and 71.0% on patching them. These are Google's reported results.
- Access starts with a subset of cyber defenders. Anthropic separately reports a research model producing working exploits with little human involvement.
- Ask whether your team can review and ship fixes as fast as automated tools can find flaws.
A flaw matters once someone proves it works
A software flaw is a rumour until someone shows it can be used. Proving that has long needed scarce specialists. Google's announcement of Gemini 4 Carbon puts that scarcity in question. The company reports a single model scored on every step: finding the flaw, proving it, and writing the fix.
The lesson here is that security work is becoming a pipeline, not a craft. When machines do the first steps quickly, the slow step is the human one: deciding what to fix, and getting the fix shipped.
What Google reports
Google announced Carbon on Wednesday, two weeks after its predecessor, Gemini 4 Argon. It says Carbon scores 88.2% on SEC-bench Pro, a test of finding vulnerabilities. It scores 54.8% on ExploitGym, which tests exploitation. It scores 71.0% on Collinear's CWE-bench, which tests patching.
Google also says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. The Hacker News reports the comparison scores: Anthropic's Claude Opus 5.5 at 58, OpenAI's GPT-6 Astra at 57 and GPT-6.1 Sol at 55.
On Google's internal tests, Carbon more than doubled Argon's success rate at writing secure code. Google says the Chrome team found Carbon more than 10 times better than another lab's frontier model at producing high-quality patches.
How the chain works
Each benchmark maps to one stage of a defender's day. Finding means locating a real weakness in real software. Exploitation means checking whether the weakness can actually be used. That check tells a team which flaws to fix first, or whether to fix a given one at all.
Patching means repairing the root cause, not hiding the symptom. Google says Carbon can run for days on such work. Its teams used it to trace and fix subtle bugs across large codebases and low-level systems.
Google notes that its external benchmark results were evaluated without guardrails. Those are the safety limits a customer would normally get. The scores show the model's raw ability, not what a product will allow.
The same skill serves both sides
Skill at proving exploits is useful to attackers and defenders alike. Anthropic disclosed this week that an unreleased research model found serious flaws in widely used software. The Hacker News lists a one-click remote code execution flaw in Chrome, a one-click exploit chain in Safari and an Android lock screen bypass. It also lists a denial-of-service flaw in the Windows HTTP.sys driver, CVE-2026-33096, with a CVSS score of 7.5.
Google, Apple and Microsoft have fixed all of them, according to the report. Anthropic said producing such chains was once the domain of a few specialist teams. It says models can now do it with little human involvement and prove it with a working exploit anyone can run.
What the numbers do not show
Most of these figures are Google's own. Only the Artificial Analysis ranking comes from a third party. The internal results have not been independently reproduced in anything provided here.
Access is also limited. Google says Carbon goes first to a subset of cyber defenders, then to more trusted partners through its Fairwind Program. It says it is taking part in the U.S. government's voluntary pre-release testing process. A date for developers and businesses is not given.
Pricing is stated. The introductory rate is $2 per million input tokens and $10 per million output tokens. After the introductory period, it rises to $4 and $20.
Where the pressure lands
If finding and proving flaws gets cheaper, the number of confirmed problems on a security team's desk can grow. Fixes still need review, testing, scheduling and sign-off. Google's own claim about patching suggests the same tools may help there too. A faster finder is still only useful to a team that can absorb what it finds.
Questions to put to your team
Ask how long a confirmed flaw takes to reach production today, and where it waits. Ask who approves an AI-written patch, and what evidence that person needs.
Ask your software vendors whether they use such tools on their own code. Ask your security provider whether it is in the early-access group, and what it will do with the findings.
Finding flaws is becoming easier to buy. Fixing them still depends on how well your organisation is run.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Google.
