Google's strongest AI model goes first to cyber-defence teams only
Google unveiled Gemini 4 Carbon on 28 October; it leads several security benchmarks and is for now available only to selected defenders in the Fairwind Programme.
According to Google, Gemini 4 Carbon solves 54.8 per cent of tasks in ExploitGym, a test that requires exploiting vulnerabilities, against 42.0 per cent for its predecessor, Argon. Google is therefore releasing the model first to vetted defence teams and has given no date for companies and developers.
What happened
The leading AI labs are moving to stagger the release of their most capable models because of their offensive potential. Google says Carbon finds a vulnerability in 88.2 per cent of cases on SEC-bench Pro and fixes 71.0 per cent on CWE-bench. Google ran these tests without guardrails, so the restricted version may score differently. According to WinBuzzer the Fairwind Programme has more than 650 partners, including government cyber agencies and critical-infrastructure operators; members need background checks and phishing-resistant authentication and may not pass on access. For Argon, broad release followed the Fairwind launch within about two weeks. On the independent Artificial Analysis index Carbon scores 59, narrowly ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). Its rate of false factual claims there rose from 4 to 11 per cent. Google has published no technical report. The model is going through the American government's voluntary pre-release testing.
What follows
Assessment: defenders gain a head start for as long as attackers lack comparable tools. How long that window lasts is unknown. For companies the main consequence is more work: if models find vulnerabilities in volume, the number of security advisories and updates to install rises, as with the recent PHP flaws. Second, access to the best models becomes a matter of vetting, not just price. Those outside the programme wait. Third, after the introductory period the price rises from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. The higher rate of factual errors argues for keeping humans in the loop.
What changes in practice
- IT departments
- If AI models find flaws faster, more security updates follow. Installing them promptly remains the main protection.
- Critical infrastructure
- Operators and agencies can gain early access through the Fairwind Programme if they pass vetting.
- Companies
- There is no date for general access. After the introductory period, per-token prices double.
What to watch
- Date of the broad release of Gemini 4 Carbon
- Number of security advisories vendors attribute to AI findings
- European agencies and infrastructure operators in the Fairwind Programme
Sources
- Google: Introducing Gemini 4 Carbon (28.10.2026)
- VentureBeat: Google unveils Gemini 4 Carbon, but access is limited
- WinBuzzer: Gemini 4 Carbon puts Google back on top, but only cyber defenders can use it