The cyber-focused model arrives two weeks after Gemini 4 Argon, with access limited to security defenders
TF investigations learned Google unveiled Gemini 4 Carbon, a flaw-finding AI model most users cannot reach. The company announced the model on 28 October 2026. Selected security teams get access within days, and nobody else has a date.
What’s Happening & Why It Matters
What Carbon Does
Google (@Google) built Carbon to spot software weaknesses, test whether attackers can exploit them, and patch the root cause. The model runs multistep work lasting days. Google teams report long runs which diagnose and fix bugs across large codebases.
Inside Google, Carbon more than doubled Argon’s secure-code success rate, the model released two weeks earlier. The Chrome team says Carbon produced high-quality security patches at over ten times the rate of an unnamed rival model. Independent testers have not verified either claim.
The Benchmark Numbers
Google reports 88.2% on SEC-bench Pro, a test of finding flaws in real software. Carbon scores 71.0% on CWE-bench, which checks whether a patch fixes the weakness. Carbon reaches 54.8% on ExploitGym, which tests controlled exploitation. Google ran all three cyber tests without guardrails, so the scores show raw ability without safety layers.
Carbon ranks first on the Artificial Analysis Intelligence Index 5.0 with 59 points, and fourth on WebDev Arena at about 1,768. Android Headlines reports Google claiming wins over GPT-6 Astra from OpenAI and Claude Opus 5.5 from Anthropic. The claims cover coding and science tasks.
Why Access Is Restricted
A model which finds, confirms and patches flaws helps attackers as much as defenders. Google ties the staged release to the dual-use risk. Selected cyber defenders go first, and more trusted partners join through the Fairwind programme. Google takes part in the US government’s voluntary pre-release testing.
Android Headlines expects developers to wait one to two more weeks, though Google set no firm date. Introductory pricing matches Argon at $2 (about €1.78) per million input tokens and $10 (about €8.90) per million output tokens. Cached input gets a 95% discount.
Safety Claims and Their Limits
Google says unsupported claims fell by about half against Argon. High-severity misalignment alerts in live coding work dropped by close to 80%. Gray Swan prompt-injection tests found no successful attack, against 0.7% for Argon. Google adds cheating attempts on Terminal-Bench 4.0 ran at 0.4%, with no success.
All the figures come from internal and vendor tests. A clean result on the tested attacks doesn’t prove immunity to the next one. TF covered another AI-driven security story in PHP2Shell: PHP Vulnerabilities, AI Research and Patches.
My take: Google is right to hold Carbon back. A model which can validate exploits is a policy question before a product question. The gap to watch is time. Defenders get a head start of days, and rivals may close the gap within months. Governments should treat the staged release as a template for cyber models.
TF Summary: What’s Next
Developers should get wider access within weeks if Google keeps the pace set in launch coverage. Security researchers will test the benchmark claims and hunt for misuse. Rivals will answer with cyber-focused models.
MY FORECAST: Gemini 4 Carbon will reach public API access by the end of November 2026. At least one major rival will announce a restricted cyber model by February 2027. Independent testers will score Carbon below Google’s ExploitGym result.
Related Stories
- PHP2Shell: PHP Vulnerabilities, AI Research and Patches
- ChatGPT Adds Locked Chats and Pins as GPT-6 Astra Nears
- Google DeepMind Staff Resign Over AI Military Contracts
- Google Rolls Out Gemini Personal Intelligence to Android Auto
- Anthropic’s IPO Filing Warns of Bioweapons Misuse, Model Blackmail and Risks to Humanity