Skip to content
Sign in
Technology · Global

Google limits Gemini 4 Carbon to selected cybersecurity firms

Google released Gemini 4 Carbon, its latest frontier model, and said it improves on Gemini 4 Argon in coding, agent tasks and cybersecurity. Google has made Carbon available to a select group of cybersecurity defense companies, while reporting also highlights benchmark results for coding and security tasks.

AI-assisted synthesis · Based on 4 independent reports from 5 outlets · Article text read from 3 of 4 reportsHow it's made

View timeline · 8 entries
5 outlets3 regionsLatest reporting 4 min read
Coverage: Center OnlyL 0 · C 3 · R 0 · 2 unrated
Updat3 graphic: Google
Updat3 graphic

What happened

Key facts

  • Google released Gemini 4 Carbon as its latest frontier model (per The Verge, aa.com.tr, the-decoder.com).
  • Google said Gemini 4 Carbon improves on Gemini 4 Argon in coding, agentic tasks and cybersecurity (per The Verge, aa.com.tr, the-decoder.com).
  • Google said it plans to expand Gemini 4 Carbon's availability beyond the selected cybersecurity companies, without specifying the scope (per The Verge, aa.com.tr, the-decoder.com).
  • Google said it improved Carbon across full-stack web development, scientific tasks and multimodal understanding based on user feedback from its previous release (per aa.com.tr, The Verge, the-decoder.com).
  • Google said Carbon rose 89 points to fourth place on the WebDev Arena leaderboard (per aa.com.tr).
  • Google reported Carbon scores of 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71% on CWE-bench for patching (per aa.com.tr).
  • Google said Carbon reduced the attack success rate on Gray Swan's indirect prompt injection benchmark from 0.7% for Argon to 0.0%, blocking every manipulation attempt tested (per the-decoder.com, The Verge, aa.com.tr).

The brief

Google has released Gemini 4 Carbon, its latest frontier artificial intelligence model, and restricted its initial availability to a select group of cybersecurity defense companies.

Google says the model improves on Gemini 4 Argon in coding, agentic tasks and cybersecurity; it also says it plans to make Carbon available beyond the initial group, without defining the scope of that expansion. The limited rollout places cybersecurity at the center of the launch, while the reported evaluations cover several different tasks.

Google described Carbon as having gains in coding, autonomous agent work and cybersecurity defense. The Verge reported that the company presented it as better at finding vulnerabilities and generating secure code, and said its internal teams had used it for complex engineering tasks spanning multiple steps and dead ends.

Google cited security benchmark results through aa.com.tr, which reported the company’s figures of 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71% on CWE-bench for patching. Those are company-reported scores on named evaluations.

The Decoder separately reported Google’s claim that Carbon achieved a 0.0% attack success rate on Gray Swan’s indirect prompt injection benchmark, down from 0.7% for Argon, with every manipulation attempt in that test blocked. The outlets also emphasized different measures of model performance.

The Verge highlighted Carbon’s position on the Artificial Analysis Intelligence Index v5 and said Claude Opus 5.5 and GPT-6 Astra trailed it. The Decoder supplied the score comparison: Carbon received 59 points, against Argon’s 50. It also reported that Claude Opus 5.5 and GPT-6 Astra each scored 50% on the index, at their maximum settings.

These index results concern a separate evaluation from the cybersecurity tests. aa.com.tr focused on additional development and benchmark claims: Google said feedback on its previous release informed improvements in full-stack web development, scientific tasks and multimodal understanding.

The outlet reported Google’s claim that Carbon gained 89 points to reach fourth place on the WebDev Arena leaderboard. It also described Carbon as Google’s strongest frontier model, with advanced cybersecurity and long-horizon capabilities; these characterizations are the company’s claims, not independently established findings in the reporting.

For coding agents, The Decoder reported a 59-point Carbon score, up from Argon’s 50 on Artificial Analysis’s Coding Agent Index, while Claude Opus 5.5 scored 62. That comparison differs from the Intelligence Index ranking: a model’s position on one evaluation does not establish its standing on another.

Across the accounts, the announcement combines a restricted cybersecurity rollout with broader claims about coding, reasoning and security performance; the available rollout detail is that Google intends eventual expansion, but has not specified its extent.

Why it matters

  • Cybersecurity defense companies in the initial selected group are the only users identified as having access to Carbon now, giving them access to a model Google says improves vulnerability discovery and secure-code generation.
  • Security teams evaluating the model have company-reported results across distinct tests: 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71% on CWE-bench.
  • Google’s reported Gray Swan result is a reduction from Argon’s 0.7% attack success rate to Carbon’s 0.0% on the tested prompt-injection benchmark; the claim concerns that evaluation, not every possible manipulation attempt.
  • For coding-agent performance, the reported index scores place Carbon at 59, below Claude Opus 5.5 at 62, while Carbon scored 59 against Argon’s 50 on the Intelligence Index v5.

What to watch next

  • Whether Google expands Carbon beyond the selected cybersecurity defense companies, and what scope it sets for that expansion.
  • Whether Google provides further detail on access to Carbon for cybersecurity defense companies.
  • How Carbon performs on the named security evaluations, including SEC-bench Pro, ExploitGym, CWE-bench and Gray Swan’s indirect prompt injection benchmark.
  • Whether subsequent Artificial Analysis Intelligence Index v5 or Coding Agent Index results change the reported comparisons with Argon and Claude Opus 5.5.

How it's being covered

Coverage analysisShow details
5 source articles · 5 unique outlets · 4 distinct reports
Outlet political mix: Center Only· 3 Center · 2 unrated
Where the outlets are based: US/Canada 40% · Middle East 20% · Other 40%
Outlet political mix
Center Only
Left
Center
Right
0 outlets
3 outlets
0 outlets
The marker shows coverage balance: each unique outlet counts once. 2 unrated outlets are not included.
Geography
Where the outlets covering this story are based.
5 unique outletsDominant coverage: US/Canada
Coverage imbalance
  • Coverage is currently concentrated among Center-rated outlets (3 of 3 rated sources); comparatively few Left or Right-rated outlets have published on the event.
How we measure coverage

How outlets' accounts compare

2 framing differences · 2 uncorroborated claims · 2 information gaps
The reports repeat the same account with different emphasis; its central claims are not independently confirmed in the reporting Updat3 read.Compared across 3 reports whose article text Updat3 read; 1 other outlet was checked by headline only.
Framing / emphasis differences
Sources emphasise different angles, actors or consequences of the same reporting.
  • The Verge foregrounded the restricted availability to cybersecurity defense companies and the model’s cybersecurity and agentic-task gains; aa.com.tr emphasized Google’s description of Carbon as its strongest frontier model and gave several benchmark scores.
  • The Decoder emphasized index scores and the Gray Swan prompt-injection result, while aa.com.tr highlighted WebDev Arena, SEC-bench Pro, ExploitGym and CWE-bench.
Updat3: These are differences of emphasis, not disputes about what is reported. They do not show that the reported claims are true.
One-sided or uncorroborated claims
Claims that come from one party or one source family and are not independently corroborated.
  • Google described Carbon as its strongest frontier model, according to aa.com.tr.
  • Google said Carbon blocked every manipulation attempt tested on Gray Swan’s indirect prompt injection benchmark, according to The Decoder.
Updat3: Not independently corroborated yet — treat as a claim, not an established fact.
Not established by the reporting
Information the available reporting does not contain or confirm. This is a gap, not a disagreement.
  • The coverage does not identify the selected cybersecurity companies or provide details on how Google chose them.
  • Google’s benchmark figures and performance characterizations are reported as company claims; the accounts do not provide independent verification of those results.
Updat3: Updat3 cannot conclude anything about this from the current sources.

Timeline

Developments from this story's sources, dated background we compiled and fact-checked, and our earlier coverage. Each row says which.
  1. Google announced Project Zero, a team that investigates vulnerabilities in software.
    Background · compiled and fact-checked by Updat3
  2. Google launched OSS-Fuzz, an automated testing service for finding bugs in open-source projects.
    Background · compiled and fact-checked by Updat3
  3. The National Institute of Standards and Technology published its AI Risk Management Framework on January 26, 2023, offering organizations a voluntary structure for identifying and managing AI risks.
    Background · from our historical context
  4. On June 8, 2023, Google announced its Secure AI Framework, a set of principles for securing AI systems and their use.
    Background · from our historical context
  5. In October 2023, Google said it was expanding its vulnerability-research work using artificial intelligence, including experiments intended to identify software flaws.
    Background · from our historical context
  6. The two names belong to Google’s Gemini model family, introduced on December 6, 2023, when the company presented Gemini as a multimodal model designed to work with text, images, audio, video and code.
    Background · from our historical context
  7. Google announced Gemini 1.5 on February 15, 2024, describing advances in long-context processing; that release illustrated the company’s continuing effort to improve general-purpose model capabilities.
    Background · from our historical context
  8. Google said its Big Sleep AI agent had found a previously unknown vulnerability in SQLite.
    Background · compiled and fact-checked by Updat3

How we got here

Before Gemini 4 Carbon, Google’s Gemini 4 Argon served as the earlier model against which Google compared the newer system. The two names belong to Google’s Gemini model family, introduced on December 6, 2023, when the company presented Gemini as a multimodal model designed to work with text, images, audio, video and code.

Google announced Gemini 1.5 on February 15, 2024, describing advances in long-context processing; that release illustrated the company’s continuing effort to improve general-purpose model capabilities. Google had built security work around its software and AI systems before the Gemini releases.

Original sources

1 of 5 articles · Unknown · Show all

Every story links to the reporting it was built from. How we choose sources

Spotted a factual error, a broken link or a misleading summary? Tell us.