Google launches Gemini 4 Carbon, a frontier model it says improves automated vulnerability detection
Google released Gemini 4 Carbon (28 Oct 2026), a model it says leads at finding software vulnerabilities, via its limited Fairwind program. Not a breach.
- Published
- Updated
- Ecosystem
- AI developer toolsCloud
- Fixed in
- N/A
- Exploitation
- N/A
What happened
- 28 October 2026: Google DeepMind announced Gemini 4 Carbon, which it calls “our strongest model with state-of-the-art cybersecurity and long horizon capabilities,” attributed to Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google. The post says it follows Gemini 4 Argon, which launched two weeks earlier (Google).
- Google says Carbon shows state-of-the-art performance across vulnerability detection, validation and remediation: it uncovers complex flaws in real-world software, checks which vulnerabilities are exploitable, and turns findings into code fixes (Google).
- Company-reported benchmarks, which Google says were evaluated for the cybersecurity tasks without guardrails: SEC-bench Pro (finding vulnerabilities) 88.2%; ExploitGym (exploitation) 54.8%; CWE-bench (patching root causes) 71.0%; and the Artificial Analysis Intelligence Index v5.0 at 59, ranked first (Google). The Hacker News reports that Index score of 59 as ahead of Claude Opus 5.5 (58), GPT-6 Astra (57) and GPT-6.1 Sol (55) (The Hacker News).
- Google reports its internal benchmarks show Carbon with a more than 2x higher success rate at generating secure code than Argon (Google). The Hacker News states the same, that Carbon is twice as likely to produce secure code as Argon (The Hacker News).
- On the Gray Swan Indirect Prompt Injection benchmark, Google reports a 0.0% attack success rate for Carbon versus 0.7% for Argon, and says it resisted every evaluated attack (Google; The Hacker News).
- Google says that on ambiguous or high-pressure tasks, Carbon surfaces discrepancies to the user and asks for human input rather than guessing (Google; The Hacker News).
- Access: it began the week of release for a subset of cyber defenders, with gradual expansion to more trusted partners through the Fairwind Program. Google says it is “actively engaged in the U.S. government’s voluntary process for pre-release model testing,” and that broader availability to developers, businesses and consumers is planned “as soon as possible,” with no date given (Google).
- Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% off, rising after the introductory period to $4 per million input and $20 per million output (Google).
About the Fairwind Program
The Fairwind Program, launched 2 September 2026, is what Google calls a “limited access program for governments and trusted partners to use our most advanced cyber defense capabilities,” giving defenders access to Gemini models to find and fix vulnerabilities autonomously (Google). Initial access is staged to governments and national cyber authorities, critical-infrastructure operators, and core technology platforms, and participants agree to operational standards such as limiting access to internal cybersecurity, incident-response or penetration-testing staff and deploying multi-factor authentication (Google). Google reports more than 650 participating partners worldwide (Google).
Why it matters
Automated vulnerability finding is dual-use: a model that is better at discovering and validating exploitable flaws helps defenders triage and patch, but the same capability lowers the cost of offensive research. Google frames its staged, defender-first rollout as giving defenders “a vital adaptation window” to harden systems before attackers gain comparable capabilities (Google). Because the headline numbers are Google’s own benchmarks, run without guardrails for the cyber tasks, they describe the model’s ceiling under test conditions rather than guaranteed results in your environment (Google).
What to do
- Nothing to remediate. There is no vulnerability, exposure or compromise here; no action is required to secure your systems against this release.
- Watch availability. Carbon is currently limited to a subset of cyber defenders through the Fairwind Program, with wider access planned but undated. If you want to use it, track the Fairwind Program and the Gemini model announcements (Google).
- Treat benchmark claims as vendor-reported. SEC-bench Pro, ExploitGym, CWE-bench and the secure-code and prompt-injection figures are Google’s own; validate against your own workloads before relying on them.
- Plan for faster offensive research. Assume capabilities like these become available to adversaries over time, and prioritise shortening your own detect-and-patch cycle.