Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Gemini 4 Carbon scores 88.2% on vulnerability discovery, but Google isn't opening it to everyone

Google has released Gemini 4 Carbon, a model that hunts for holes in code and can work on a complex task for days without human input. On SEC-bench Pro, a vulnerability-discovery test, it scored 88.2%. On ExploitGym, where the flaws must also be exploited, it scored 54.8%. But only cyber defenders and trusted partners will get access, even though the model ranks first on the Artificial Analysis leaderboard. Everyone else will have to wait for a public release and independent verification.

AI-processed from TG: @ai_product (AI Product / Igor Akimov); edited by Hamidun News
Gemini 4 Carbon scores 88.2% on vulnerability discovery, but Google isn't opening it to everyone
Source: TG: @ai_product (AI Product / Igor Akimov). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @ai_product (AI Product / Igor Akimov) ↗
Editor-in-Chief
Key metric #1CWE-bench, patch preparation for vulnerabilities: 71.0%
Key metric #2Produces twice as much secure code as Argon
WebDev Arena4th place, rating 1769 (+89 Elo)

Picture an assistant that reads a huge repository on its own, finds dangerous spots in it, checks whether they can actually be exploited, and prepares a fix. This is the role Google sees for its new model, Gemini 4 Carbon, which arrived just two weeks after its predecessor, Argon. On SEC-bench Pro, a test where the model searches for vulnerabilities in code, it scored 88.2%. And on the independent Artificial Analysis Intelligence Index v5.0, which condenses model results into a single overall score, Carbon immediately took first place. There's a catch: the version that leads this index won't go to all developers, only to cyber defenders and selected partners.

What Carbon can do and the numbers Google published

The announcement focuses on cybersecurity and long agentic tasks. An agentic task is one where a model doesn't answer a single question but carries out a chain of steps on its own. According to Google, Carbon can work on a complex task for days. For defenders the scenario is clear: scan a codebase, find weak spots, confirm them, and propose a patch.

Here are the results from the announcement:

  • SEC-bench Pro, vulnerability discovery: 88.2%.
  • ExploitGym, exploitation of discovered flaws, meaning a check that they can really be used: 54.8%.
  • CWE-bench, patch preparation for common vulnerability classes: 71.0%.
  • WebDev Arena, a competition among models for building web applications: 4th place, rating 1769, up 89 Elo points.

Google also states separately that Carbon generates twice as much secure code as Argon. In practice, this means the model is less likely to leave typical flaws in a project, such as unvalidated user input.

The Chrome team goes further and claims that, in patch quality, Carbon is 10 times better than the "large model from another lab." The competitor's name is not disclosed. This is Google's own assessment, so it should be treated as a developer's claim, not an independent measurement.

Why access was restricted

Only cyber defenders and "trusted partners" under the Fairwind Program will get the version that leads the index. In addition, Carbon is undergoing voluntary testing by the US government.

The logic is clear. If a model finds vulnerabilities, and in 54.8% of ExploitGym cases it also successfully exploits them, then in the wrong hands it becomes an attack tool. So it goes first to those who fix things, and only later will decisions be made about a broad launch. Google is not the first to choose this approach, but here it is applied to a model that simultaneously leads the overall ranking.

What to do in practice

If you are responsible for product security, it's worth learning the terms of the Fairwind Program in advance. For teams with large codebases, automated vulnerability discovery and patching could take a noticeable share of routine work off their plates.

Don't expect the public Gemini to immediately gain the same capabilities. The announcement doesn't say when, or in what form, Carbon will become available to everyone, or which restrictions will remain.

Benchmarks call for caution. The Artificial Analysis leaderboard is independent, but the SEC-bench Pro, ExploitGym and CWE-bench figures, as well as the patch comparison with a competitor's model, come from Google itself. There is no third-party confirmation yet: no one outside can verify the model.

Finally, there is pace. The two weeks between Argon and Carbon show how quickly Google is updating its lineup. Teams embedding models into products are better off not locking in to a single version and keeping an abstraction layer so they can swap models without rewriting their pipeline.

For now, Carbon is first and foremost Google's bid for leadership in AI for cybersecurity. Only selected partners can test it, and everyone else must wait for a public release and independent tests.

Story

Key Questions Answered

What Carbon can do and the numbers Google published?

The announcement focuses on cybersecurity and long agentic tasks. An agentic task is one where a model doesn't answer a single question but carries out a chain of steps on its own. According to Google, Carbon can work on a complex task for days. For defenders the scenario is clear: scan a codebase, find weak spots, confirm them, and propose a patch. Here are the results from the announcement: SEC-bench Pro, vulnerability discovery: 88.2%.

Why access was restricted?

Only cyber defenders and "trusted partners" under the Fairwind Program will get the version that leads the index. In addition, Carbon is undergoing voluntary testing by the US government. The logic is clear. If a model finds vulnerabilities, and in 54.8% of ExploitGym cases it also successfully exploits them, then in the wrong hands it becomes an attack tool. So it goes first to those who fix things, and only later will decisions be made about a broad launch.

What to do in practice?

If you are responsible for product security, it's worth learning the terms of the Fairwind Program in advance. For teams with large codebases, automated vulnerability discovery and patching could take a noticeable share of routine work off their plates. Don't expect the public Gemini to immediately gain the same capabilities. The announcement doesn't say when, or in what form, Carbon will become available to everyone, or which restrictions will remain. Benchmarks call for caution.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…