Google announced Gemini 4 Carbon on 28 October, promising to give selected cyber defenders access first. The company describes a general frontier model with stronger cybersecurity and long-horizon capabilities, with wider developer, business and consumer availability to follow.

Google said it would start making Carbon available to a subset of defenders that week, then gradually extend access to additional trusted partners through its Fairwind Program. As of 30 October, the reviewed public sources did not establish which organizations had received access, which regions were supported or a public Carbon API endpoint.

There is outside evidence for part of the performance story. Artificial Analysis independently places Carbon at the top of its general Intelligence Index, and Collinear’s cybersecurity leaderboard lists a 71% patching result under defined conditions. Google’s vulnerability-finding, exploitation and prompt-injection figures remain vendor-reported claims.

What defenders can learn from Fairwind’s access policy

The Fairwind Program page describes priorities and obligations that help explain Google’s defender-first approach. Priority groups include governments and national authorities, critical infrastructure organizations in sectors such as healthcare, telecommunications, energy and finance, and core technology providers. Academic labs working on defensive benchmarks can also apply.

Applicants are directed to their Google Cloud account team. The program describes background checks covering ethical and security records, without guaranteeing a response time. Its security requirements include individual authentication, phishing-resistant multifactor authentication, access controls and tracking employee use. Access is intended for internal cybersecurity, incident-response and penetration-testing teams; sharing, resale and redistribution are prohibited.

Permitted work includes authorized attack simulation, reverse engineering and malware analysis for defense or academic research. Malicious malware creation is prohibited. These distinctions matter to a security team assessing whether its work fits the program, especially when defensive investigations involve techniques also used by attackers.

There is a documentation gap: Fairwind’s technology descriptions and FAQ still explicitly name Gemini 4 Argon. The page provides program guidance, but it does not establish Carbon-specific integrations, data-retention terms or delivery arrangements. Its descriptions of standalone access, CodeMender and managed Gemini Enterprise should not automatically be carried over to Carbon.

Google announced introductory Carbon pricing of $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95%. The announcement’s footnote lists subsequent rates of $4 and $20, without specifying the introductory period’s duration. These are announced prices; they do not establish that a public API is currently available at those rates. No Carbon weights were announced in the reviewed sources.

Artificial Analysis corroborates a general score of 59

The clearest independent measurement is the Artificial Analysis Intelligence Index v5.0. Its page says Artificial Analysis conducts the evaluations independently and lists Gemini 4 Carbon in High reasoning mode at 59, ahead of the next listed result at 58, when checked on 30 October.

The index combines nine evaluations of knowledge work, coding, science and general knowledge into a composite score. Its tasks include terminal work and demanding reasoning questions.

This supports a specific conclusion: Carbon performs strongly across the tasks combined in that index under Artificial Analysis’s evaluation setup. A one-point lead does not establish meaningful superiority for every workload, and the index alone cannot tell a security team how reliably Carbon will audit its repositories or resist malicious instructions in its tools.

The 71% patching score belongs to a particular agent setup

Collinear’s CWE-bench leaderboard lists Carbon running in Antigravity at 71% on its programmatic pass@1 measure. The suite contains 120 private held-out audit-and-patch tasks. Agents receive a source checkout and an audit instruction, without being told the flaw’s location or how many vulnerabilities it contains.

The evaluation uses high reasoning, four rollouts per task and a one-hour limit for each rollout. Programmatic success requires the patch to block the exploit while preserving the repository’s existing tests. A separate panel of model judges assesses criteria including generalization and integrity.

Those conditions make the result more informative than an isolated percentage. The task tests whether an agent can investigate unfamiliar code and repair it while keeping existing functionality working. Its score reflects the model, agent harness and available time together. The leaderboard’s programmatic pass@4 result rises to 81%, showing that the reported outcome also depends on the number of attempts allowed.

The provenance needs care. Collinear’s benchmark explanation distinguishes its leaderboard from Artificial Analysis’s separate CWE-Bench-AA implementation and says the scores are not directly comparable. Artificial Analysis’s methodology describes a different harness, one run per task and a two-hour budget. The 71% listed by Collinear is not an independently replicated Artificial Analysis result; the operator of that exact run remains unresolved in the reviewed evidence.

Collinear’s listed 71% measures the share of successful audit-and-patch attempts on this suite in the Antigravity setup, with roughly 29% failing.

Google’s finding and exploitation scores need fuller methods

Google reports 88.2% on SEC-bench Pro for vulnerability finding and 54.8% on ExploitGym for exploitation. Its announcement explicitly says the cybersecurity benchmark results were evaluated without guardrails.

The reviewed evidence did not establish the operator, model checkpoint, sample, prompts, attempt counts or budgets for those two results. Google’s model performance page repeats the figures, but another Google page is the same vendor origin. The linked Carbon methodology document was not accessible in the evidence gathered for this article.

The scores therefore remain Google’s reported results. They offer a reason to investigate Carbon’s capabilities, while leaving important questions about reproduction and comparison open.

The exploitation figure also explains the policy stakes. An ability to identify a weakness can help a defender repair software; an ability to exploit it can help establish impact and can also enable misuse. Fairwind’s vetting, contractual restrictions and access controls govern who may use the system and for what purposes. They do not remove those underlying dual-use capabilities. The unguarded benchmark conditions also leave open how the measured capabilities will translate into the restricted service.

A reported 0.0% injection success rate has a limited scope

Google says Carbon recorded a 0.0% attack success rate on Gray Swan’s indirect prompt-injection evaluation. Indirect injection involves instructions arriving through material an agent encounters, such as a document, web page or tool output, rather than through its authorized user’s request.

Gray Swan describes evaluations covering tool use, coding agents and browser agents. The reviewed material did not independently publish Carbon’s exact result or establish its sample size, attack budget, adaptation process, scoring or deployment configuration.

A reported zero is an observation within an evaluation. Without those details, readers cannot judge how extensively attackers searched for failures or how closely the tested environment resembles a proposed deployment. It does not establish immunity to future attacks.

Google describes Carbon as building on a layered safety stack. Its earlier Argon safety account includes refusal training, input screening, activation probes and continuous monitoring, with possible intervention during a run. These are complementary controls. Their presence does not independently verify Carbon’s resistance across tools and operating environments.

For a team considering access, the useful next questions are concrete: which Carbon configuration will it receive, what tools may it use, what data-handling terms apply, and what evaluation evidence matches that deployment?