Google launches Gemini 4 Carbon, a frontier model built for cyber defence
Google DeepMind released Gemini 4 Carbon on 28 October 2026, scoring 88.2% on the SEC-bench Pro vulnerability-finding benchmark and rolling out first to cyber defenders, not the open market.
Google DeepMind introduced Gemini 4 Carbon on 28 October 2026, calling it "our strongest model with state-of-the-art cybersecurity and long horizon capabilities," according to the company's blog post. Google took the wraps off the model on Wednesday, giving a subset of cyber defenders access before anyone else, The Hacker News reported on 29 October 2026.
Benchmarks, pricing and rollout
| Benchmark | Score | What it measures |
|---|---|---|
| SEC-bench Pro | 88.2% | finding real vulnerabilities |
| ExploitGym | 54.8% | validating which flaws are exploitable |
| CWE-bench | 71.0% | patching the root cause |
- Carbon ranks first on the updated version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, ahead of Anthropic's Claude Opus 5.5 (58) and OpenAI's GPT-6 Astra (57), The Hacker News reported.
- Internally, Carbon generates secure code at more than twice the success rate of its predecessor, Gemini 4 Argon, and the Chrome team found it produced patches "more than 10X better" than a rival lab's large frontier model, Google said.
- The model recorded a 0.0% attack success rate on Gray Swan's Indirect Prompt Injection Benchmark, improving on Argon's 0.7%, according to Google.
- Carbon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period, with cached input tokens discounted 95%, Google said.
- Access starts this week with "a subset of cyber defenders" and expands gradually through Google's Fairwind Program before reaching developers, businesses and consumers, the company said.
Why it matters
Frontier labs are now racing on who can automate vulnerability discovery and patching fastest, and Google's decision to give Carbon first to defenders rather than the open market shows how seriously it treats the risk that the same offensive-capable coding skills could be turned against poorly secured systems. For companies in the Caspian region deploying agentic AI tools on live infrastructure, the staged rollout — and Google's own description of a model that autonomously finds, validates and patches exploits — is a reminder that security access controls need to keep pace with what these models can already do once broader release follows.
