Gemini 4 Is Here, and Google’s Flagship Tops All Other AI Models on Cybersecurity

What it is

Google's Gemini 4 Argon, its frontier model, positioned specifically for cyber defense. The catalog summary already carries the headline security numbers (0.7% prompt-injection attack success, 70.9% Wiz benchmark, Fairwind, "without cyber guardrails"). What follows is what those lines cannot carry: the access model, the baselines, and the caveats an engineer needs before treating any of it as fact.

How it works

The 70.9% Wiz Penetration Test Benchmark figure is a jump from 58.2% — that baseline is Gemini 3.8 Flash Cyber, the earlier restricted model Google shipped through Fairwind. So the "70.9%" is a generational delta over a prior cyber-tuned model, not over a general one. The Wiz test asks the model to write working exploits against real web-app flaws without seeing the code, scored on first-try solve rate.

The prompt-injection benchmark (Gray Swan) scores how often attacks succeed within 15 tries; lower is better. Grok 4.6 and Kimi K3, the two poor performers, scored 51.8% and 52.7% respectively (summary rounds both to ~52%).

Preconditions / Affected

Access is gated. Argon ships first to vetted defenders via the Fairwind Program — launched September 2, now 650+ partners including governments and critical-infrastructure operators — then to paid API customers and Google AI Ultra subscribers. Only the Fairwind build ships without cyber guardrails; the general release is not stated to lack them. Google is also routing pre-release access through the U.S. government's voluntary process.

Introductory API pricing: $2 / million input tokens, $10 / million output. Standard rates (start date unstated) are $4 / $20. Max output rose to 1M tokens (from 64K).

Caveats

• Self-scored benchmarks. Google computed its own DeepSWE score while rivals' numbers came from public leaderboards and company reports. The Wiz benchmark is also internal to Google. Only the Gray Swan injection numbers are third-party. Treat the coding and cyber leads as vendor-reported.
• Concedes ground. Even in Google's own table, Argon leads 12 of 18 benchmarks, ties 1, and trails on 5.
• Precedent, not novelty. Gated cyber models already exist: an early Claude Mythos helped find 271 Firefox vulnerabilities Mozilla patched; OpenAI runs Trusted Access for Cyber. The Wiz healthcare-flaw claim is a single vendor-attributed anecdote, unverified.