Google says Gemini 4 Carbon leads benchmarks for cyber tasks

Google has unveiled Gemini 4 Carbon, described by Google DeepMind’s Koray Kavukcuoglu as a frontier model with strong cybersecurity and long-horizon capabilities. In Google’s framing, Carbon leads several benchmarks: SEC-bench Pro scores 88.2% for finding vulnerabilities, ExploitGym 54.8% for exploitation, and CWE-bench 71.0% for patching. On the Artificial Analysis Intelligence Index, Carbon tops the field with 59, ahead of Anthropic’s Claude Opus 5.5 (58) and OpenAI’s GPT‑6 Astra (57) and GPT‑6.1 Sol (55).

Google stresses that Carbon’s cybersecurity focus includes improved vulnerability detection, validation and remediation, and notes Carbon is about twice as likely to generate secure code compared with Argon. Kavukcuoglu also said Carbon seeks human input when faced with ambiguity rather than guessing, and resists indirect prompt injection during evaluated attacks.

The rollout plan mirrors that of Gemini 4 Argon, with access initialised for a subset of cyber defenders under Google’s Fairwind Programme, before wider sharing with trusted partners. Google indicated active participation in a U.S. government voluntary pre-release model-testing process.

The piece situates these claims against OpenAI’s 6.1 Terra and Anthropic’s disclosed vulnerabilities, including a report from Anthropic about a unreleased internal model finding critical flaws in widely used software and resulting fixes by Google, Apple and Microsoft. The article references a CVE: CVE-2026-33096 (Windows HTTP.sys) noted in the context of DoS-like exploit chains fixed by vendors, and notes that OpenAI and Anthropic cited multi‑vendor vulnerability demonstrations concerning browser and OS products.

The report notes Carbon’s stated objective to improve detection, validation and remediation in cybersecurity use cases and to resist exploitation attempts, situating Carbon as a leading frontier model within Google’s cybersecurity-focused ambitions.