Google unveils Gemini 4 Carbon with leading cyber benchmarks
Original title: Google's Gemini 4 Carbon Leads Benchmarks for Finding, Exploiting, and Patching Vulnerabilities
Access is limited to some cyber defenders under Fairwind, and Google says it resisted every tested indirect prompt injection.
Google DeepMind released Gemini 4 Carbon, citing SEC-bench Pro at 88.2%, ExploitGym at 54.8% and CWE-bench at 71.0%, plus an Artificial Analysis score of 59. Google says Carbon is twice as likely as Argon to generate secure code and is in a limited defender rollout while US voluntary pre-release testing continues. The report also notes OpenAI's GPT-6.1 Terra and an Anthropic research model that found since-fixed flaws, including CVE-2026-33096.
Unconfirmed: Benchmark leads and prompt-injection resistance are Google's reported results, not independent confirmation.
- Carbon scored 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57.
- Google says Carbon asks for human input instead of guessing on ambiguous tasks.
- Anthropic said an unreleased model found a one-click remote code execution in Chrome.
- It also reported a Safari chain, an Android lock-screen bypass and CVE-2026-33096.
- Google, Apple and Microsoft have fixed the issues Anthropic described.
- CVEs
- CVE-2026-33096
- Organizations
- Google DeepMind
- Countries
- United States
Vulnerabilities mentioned
All| CVE | Vulnerability | CVSS | EPSS | Signals | Published |
|---|---|---|---|---|---|
| CVE-2026-33096 | Out-of-bounds read in Windows HTTP.sys allows an unauthorized attacker to deny service over a network | 7.5 | 1% |
From the original article
2 min read at thehackernews.com
Google on Wednesday took the wraps off Gemini 4 Carbon, its latest frontier model with state-of-the-art cybersecurity and long-horizon capabilities. "It shows significant gains across coding and agentic tasks as well as across key capabilities critical for cybersecurity defense, landing a firm lead in SEC-bench Pro (88. 2%) for finding vulnerabilities, ExploitGym (54.
thehackernews.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.