Google on Wednesday took the wraps off Gemini 4 Carbon, its newest frontier mannequin with state-of-the-art cybersecurity and long-horizon capabilities.
“It reveals vital positive factors throughout coding and agentic duties in addition to throughout key capabilities crucial for cybersecurity protection, touchdown a agency lead in SEC-bench Professional (88.2%) for locating vulnerabilities, ExploitGym (54.8%) for exploitation, and CWE-bench (71.0%) for patching,” Koray Kavukcuoglu, senior vp of Google DeepMind and chief AI architect at Google, stated.
On the Synthetic Evaluation Intelligence Index benchmark, Carbon leads with a rating of 59, surpassing Anthropic’s Claude Opus 5.5 (58), OpenAI GPT-6 Astra (57), GPT-6.1 Sol (55).
One of many core focus areas is cybersecurity, the place the newest iteration reportedly delivers vital efficiency enhancements throughout the board, together with vulnerability detection, validation, and remediation. Carbon can be twice as prone to generate safe code versus Argon.
Google additionally emphasised that Carbon makes an attempt to reply to advanced or high-pressure duties by in search of human enter within the face of ambiguity moderately than “guessing or forcing a path ahead,” in addition to resisting each evaluated assault try at oblique immediate injection.
The debut of Gemini 4 Carbon comes lower than every week after the tech large unveiled Gemini 4 Argon. The discharge plan is much like that of its predecessor, with the rollout at present restricted to a subset of cyber defenders earlier than increasing entry to different trusted companions as a part of its Fairwind Program.
Google additionally stated it is actively engaged within the U.S. authorities’s voluntary course of for pre-release mannequin testing.
The event comes as OpenAI launched GPT-6.1 Terra, which it described as a quick, environment friendly mannequin that “excels at communication and computer-based duties.”
“GPT‑6.1 Terra meets the Important threshold in cybersecurity beneath our Preparedness Framework, and maintains the protections developed for GPT‑6 Astra,” OpenAI stated.
“GPT-6.1 Terra brings with it the alignment enhancements of the GPT-6 collection. In line with our evaluations, GPT-6.1 Terra is extra prone to keep inside scope and actually disclose limitations and potential blockers to the consumer in comparison with GPT-5.6 Terra. In our prompt-injection assessments, GPT-6.1 Terra additionally confirmed robust resistance to malicious directions designed to divert it from the consumer’s process.”
Earlier this week, rival Anthropic additionally disclosed that an unreleased, inner analysis mannequin uncovered crucial vulnerabilities in among the world’s most generally used software program.
This included a one-click distant code execution in Google Chrome, a one-click exploit chain in Apple Safari that may very well be used to entry a consumer’s searching historical past, cookies, and different delicate information, an Android lock display screen bypass, and a denial-of-service (DoS) flaw within the Home windows HTTP.sys driver (CVE-2026-33096, CVSS rating: 7.5). All of the recognized points have been fastened by Google, Apple, and Microsoft.
“These assault chains show the mannequin’s skill to create subtle, multi-step exploits in each open-source and proprietary software program,” Anthropic stated.
“These abilities have traditionally been the area of few specialist groups with experience in these targets. Now, it’s one thing fashions can do with little human involvement, after which show by producing a working exploit that anybody can run and confirm.”
