Skip to content

Back to the catalog

Google Touts Powerful Cybersecurity Capabilities in Gemini 4 Carbon Model
securityweek.comarticle

Google Touts Powerful Cybersecurity Capabilities in Gemini 4 Carbon Model

Gemini 4 Carbon is a Google frontier model announced with restricted access for cyber defenders through the Fairwind Program (650+ organizations). Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, ahead of GPT-6 Astra and Claude models, with no guardrails. It also cites prompt injection and cheating-resistance results, and pricing of $2/$10 per million tokens.

TL;DR Google unveils Gemini 4 Carbon, claiming top benchmark scores for finding, exploiting and patching vulnerabilities, with staged access for approved defenders.
Also seen from another angle🧠 AI →
Document

What it is

Gemini 4 Carbon is a Google frontier model announced on Wednesday, with cyber defense as its stated focus. Google says it leads rivals on vulnerability discovery, exploitation and patching. It follows Gemini 4 Argon, launched two weeks earlier. Some Fairwind partners already have exclusive access to Argon's cyber defense capabilities.

Results Google reports

  • SEC-bench Pro (finding flaws in real-world software): 88.2%.
  • ExploitGym (exploitation): 54.8%.
  • Collinear's CWE-bench (patching): 71.0%.
  • Google says these results top OpenAI's GPT-6 Astra, Anthropic's Claude Opus 5.5 and Claude Fable 5.1, in Google's own comparison.
  • Against Argon, Carbon is more than twice as successful at writing secure code, by Google's internal benchmarks.
  • Google's Chrome team found Carbon's patches beat a large rival frontier model by more than 10 times on patch quality.
  • Wiz, which Google owns, says Carbon took the top spot on web penetration testing benchmarks.

On safety, no attack succeeded in Gray Swan's indirect prompt injection benchmark (Argon: 0.7%). On Terminal-Bench 4.0, none of Carbon's cheating attempts succeeded. Carbon made about half as many unsupported claims as Argon. Google's internal teams saw a nearly 80% drop in high-severity misalignment alerts in real coding environments.

Preconditions / Access

  • Carbon is not generally available. Access opens this week to a select group of cyber defenders, then expands in stages to approved partners through the Fairwind Program.
  • Fairwind covers governments, critical infrastructure operators and major technology platforms. Google says more than 650 organizations participate.
  • Availability to developers, businesses and consumers through Google product surfaces depends on Google gathering feedback and iterating on guardrails. No date is given.
  • Launch pricing matches Argon's introductory rates: $2 per million input tokens and $10 per million output tokens. After the introductory period the price rises to $4 and $20. The page does not say when that period ends.

Caveats

  • The benchmark results come from Google's own comparison and were obtained with no guardrails in place. They may not match what a guarded release does.
  • The Chrome and Argon comparisons are internal. The Wiz result comes from a Google-owned company.
  • Google frames exploitation as a defensive step: defenders use models to check which vulnerabilities are exploitable and so decide what to fix first.
  • Google says it is involved in the US government's voluntary pre-release model testing effort. The page gives no results from that testing.

Techniques named here

  • AI-assisted vulnerability discovery in real-world software
  • AI-generated vulnerability patching
  • Secure code writing
  • Web penetration testing with AI models
  • Indirect prompt injection resistance testing
  • Indirect prompt injection