# Google releases Gemini 4 Carbon for cybersecurity and long-running agent tasks

> Google DeepMind's Gemini 4 Carbon is built for cybersecurity and long-running agent work and is going first to selected defenders.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: Safety & Security
- Published: 2026-11-02T23:04:05+00:00
- Updated: 2026-11-02T23:04:05+00:00
- Web page: https://ai-news.ikangai.com/story/google-releases-gemini-4-carbon-for-cybersecurity-and-long-running-agent-tasks/
- Seen in: Turing Post

Google has released Gemini 4 Carbon, which the company calls its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. It is not reaching the public yet. Google will first make it available to a subset of cyber defenders, then gradually extend access to additional trusted partners through its Fairwind Program. The company says it will bring Carbon to developers, businesses and consumers only after it has gathered feedback and iterated on guardrails.

The announcement, dated October 28, 2026 and written by Koray Kavukcuoglu, Google DeepMind's SVP and Chief AI Architect, puts much of its weight on security benchmarks. Each was run without guardrails. Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. The post describes the ExploitGym result as a new state of the art. It says Carbon ranks #1 on the v5.0 Artificial Analysis Intelligence Index at 59. The post doesn't say what the other models on that index scored, so the lead is relative to a ranking whose comparison set is not shown in the post.

Google says its teams have seen Carbon run for days, persisting through dead ends while root-causing and patching subtle bugs across large codebases and low-level systems. The post also reports that Carbon lifted WebDev Arena by 89 Elo points to 1768, placing it at #4.

Several safety claims are internal results, and some are framed against Gemini 4 Argon, the previous model. Google says successful cheating on Terminal-Bench 4.0 was reduced to 0.0%, with attempts to cheat dropping to 0.4%. It says high-severity misalignment alerts in live coding environments fell by almost 80%. It also says Carbon cuts unsupported claims by about 2x over Argon and resists prompt injection in every evaluated attack. On Gray Swan's Indirect Prompt Injection benchmark, Carbon scored 0.0% against Argon's 0.7%.

The post's cybersecurity examples come from Google's own teams and products. The Chrome team reported that Carbon performed more than 10X better than another lab's large frontier model when generating high-quality patches. Wiz, Google's cloud security platform, observes Carbon leading on benchmarks for web penetration testing, a core capability for Wiz's Scan for Good initiative. The post doesn't name the other lab or give the patch-quality measure behind the 10X figure.

Pricing is set for an introductory period. Carbon will cost $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% off the input price. The post says the price rises to $4 per million input tokens and $20 per million output tokens after the introductory period expires. It doesn't say when that period ends or when general availability will begin. The post presents the rollout as a deliberate process that includes the U.S. government's voluntary pre-release model testing.

Google's post places the launch in a quick cadence. It refers to Gemini 4 Argon as launched two weeks earlier, and says Carbon builds on Argon's safety and security stack. That stack combines model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. Turing Post's newsletter also lists Gemini 4 Carbon among the week's model releases. It describes the model only as built for cybersecurity and long-running agent tasks and going first to selected defenders, with no benchmarks.

## Why it matters

Security teams may get early access to agent-focused models for defensive work, while other companies cannot yet evaluate it.

## What you can do

Security teams that qualify for early access could request evaluation to test it on their own defensive workflows.

## Original source

- [Introducing Gemini 4 Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.