# Google’s Gemini 4 Carbon is coming first to cyber defenders

Agent contract: 1.2.0

Google’s newest frontier model promises stronger coding and cybersecurity help. Access starts with selected defenders, not everyone with a Gemini subscription.

Canonical: https://brightaifuture.com/discoveries/google-gemini-4-carbon-cyber-defenders
Format: discovery
Source publication: 2026-10-28
Bright publication: 2026-10-28
Substantive update: None recorded
Evidence and review: Experimental; confidence: unassessed; approved; ai-assisted. Primary sources checked; company claims and access caveats retained. Original illustration reviewed; publication approved.

## Editorial image

An illustrated engineer studies a glowing crystal in a sunlit workshop.: https://brightaifuture.com/media/content/38a1350cb4bbb78dfc8841d52e1febe70a84e0737cd6de5d31c22a29fabe3d95.png

Credit: AI-generated editorial illustration for Bright. Conceptual, not actual Google hardware.. License: Original artwork commissioned for Bright.

Source: https://brightaifuture.com/media/content/38a1350cb4bbb78dfc8841d52e1febe70a84e0737cd6de5d31c22a29fabe3d95.png

illustration; not Actual Google hardware or demonstrated results.

## Google’s Gemini 4 Carbon is coming first to cyber defenders

Google announced Gemini 4 Carbon on October 28, with selected cyber defenders due to receive access this week. Trusted partners follow through Fairwind; broader access for developers, businesses and consumers remains ahead. [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)

Google reports stronger coding, sustained engineering and reliability. These are company findings, not independently established outcomes.

## Capability is not a finished service

Its model table lists Carbon at 71% on CWE-bench versus Argon’s 68%, and 88.2% on SEC-bench Pro versus 83.4%. These are test scores, not the share of all software flaws a model can find or fix. Carbon also trails some listed competitors on FrontierSWE 2 and OSWorld-2.1 offline. [Model comparison](https://deepmind.google/models/gemini/)

The cybersecurity benchmarks were run without guardrails. Deployed restrictions may therefore change what a model will do.

Fairwind vets partners and restricts use to defensive or research purposes. Its program page still describes Argon in places; use the Carbon announcement for the new rollout, rather than assuming every partner already has access. [Fairwind](https://deepmind.google/fairwind-program/)

For everyday people, the hopeful possibility is software that gets repaired more effectively. That benefit still needs evidence: reproducible findings, reviewed repairs and a clear record of what remains unresolved. A stronger benchmark result is not a reason to buy a subscription expecting immediate access.

## Provenance and history

{
  "dates": {
    "captureDate": "2026-10-28",
    "eventDate": "2026-10-28",
    "lastReviewedDate": "2026-10-28",
    "publicationDate": "2026-10-28"
  },
  "provenance": {
    "origin": "editorial",
    "externalId": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/"
  },
  "revisions": [
    {
      "id": "google-gemini-4-carbon-cyber-defenders-initial",
      "recordedAt": "2026-10-28",
      "sourceIds": [
        "source:google-gemini-4-carbon-cyber-defenders:0",
        "source:google-gemini-4-carbon-cyber-defenders:1",
        "source:google-gemini-4-carbon-cyber-defenders:2"
      ],
      "summary": "Original reporting on Carbon’s selective rollout."
    }
  ],
  "corrections": []
}

## Original sources

- [Google · Carbon announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [DeepMind · Gemini](https://deepmind.google/models/gemini/)
- [DeepMind · Fairwind](https://deepmind.google/fairwind-program/)

## Continue exploring

- [Sophos says AI agents bring some cyber responses down to 89 seconds](https://brightaifuture.com/discoveries/sophos-ai-cyber-response-89-seconds)