Google has made Gemini 4 Carbon, a model it says outperforms rival systems at finding, exploiting and fixing software vulnerabilities, available to paying developers and businesses, SecurityWeek reported Tuesday. A cyber-tuned version stays limited to vetted defenders, so the wider release tests whether built-in safeguards, rather than access limits, can keep the general model from helping attackers.

What Happened

In a Google blog post, Koray Kavukcuoglu, senior vice president of Google DeepMind and the company's chief AI architect, said Carbon is now offered through the paid tier of the Gemini API in Google AI Studio and Android Studio, in Gemini for business and in Google's Antigravity agent platform. It is also rolling out to Google AI Ultra subscribers in the Gemini app and AI Mode in Search.

Until now, only defenders in Fairwind, Google's early-access program for governments, critical infrastructure operators and major technology platforms, could use Carbon. More than 650 organizations take part, SecurityWeek reported Oct. 29. The variant called Carbon for cyber still requires Fairwind membership.

RelatedOpenAI Launches Agent-Run Pen Testing and Patching Tools; Brockman Ties Security to Frontier Models

At that October debut, Google said Carbon scored 88.2% on SEC-bench Pro, a test of finding flaws in real software; 54.8% on ExploitGym, which measures building working attacks; and 71.0% on CWE-bench, which measures patching. In Google's comparison, that put it ahead of OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1. Google noted the scores were obtained without guardrails.

Google cited early field results: Wiz, the cloud security company Google owns, used Carbon for cyber to find previously unknown flaws at two major software-as-a-service providers in under two hours, and Google's Cloud Vulnerability Research team built test attacks proving two bugs in a popular open-source project were exploitable, prompting maintainers to rate one critical and patch it.

Introductory API pricing is $2 per million input tokens and $10 per million output tokens through Nov. 30, rising to $4 and $20 afterward, Google said.

The Signal Proof

Google DeepMind also introduced AlphaProof Code, which pairs human experts with AI agents to formally verify software, meaning to prove mathematically that code meets a written specification. The team applied it to libsignal, the reference code for the Signal encryption protocol, versions of which also run in WhatsApp and Google Messages. Humans working with agents built on Carbon and Claude Opus 5.5 wrote the specifications; Carbon-based agents then produced about 6.5 million lines of proofs in Lean, a language whose checker rejects any invalid proof. Five researchers finished in four weeks.

The guarantees have limits. They cover Google's own build of the code, not Signal's apps, and side channels, indirect leaks such as timing, were excluded from parts of the analysis. The Register noted the proofs assume the underlying cryptography is perfect.

Why It Matters

Exploit-writing skill cuts both ways. Defenders use it to learn which bugs are truly dangerous and fix those first; attackers could use it to shrink the time between a flaw's disclosure and its abuse. Google's answer is safeguards. It says Carbon blocked 99% of attack attempts by an adaptive automated red-teaming agent in simulated environments, with a per-attempt success rate of 0.8% against 3.1% for its predecessor, Argon. Automated cyberattack requests from testing firm Gray Swan succeeded 2.17% of the time, against 9.86% for Argon, and outside tester Trajectory Labs found no new universal jailbreaks, prompts that disable safeguards across the board, in 132 hours. Google concedes that resistance to prompt injection, hidden instructions planted in content a model reads, needs more work.

Those numbers come from Google and testers it chose. For boards, exploit-building help is now a commercial product, which makes patching speed a business risk measured in days.

Context

Google followed the same order with Argon, which The Primary Wire reported went first to Fairwind defenders in September. Rivals still hold their strongest cyber models back: OpenAI is giving GPT-6 Cyber only to vetted red teams, The Primary Wire reported Nov. 5, citing Fortune, and Anthropic loosens Claude's cyber safeguards only for verified security users in a tiered program, The Primary Wire reported in October. On Monday, The Primary Wire reported that DeepMind agents cut the cost of attacking legacy 1024-bit RSA keys.

What to Watch

The Nov. 30 price change, findings from government and other outside testers, disclosure details for the open-source bugs Google's team confirmed, and whether OpenAI or Anthropic widen access to their own cyber models.