Google has introduced Gemini 4 Carbon, a new artificial intelligence model designed to identify software vulnerabilities, generate security patches, and handle complex engineering tasks that can take days to complete.
The model achieved an 88.2% score on SEC-bench Pro, a cybersecurity benchmark that tests an AI system’s ability to identify vulnerabilities in real-world software. It also scored 71% on CWE-bench for fixing security flaws and 54.8% on ExploitGym for validating whether vulnerabilities can be exploited.
Google says the results place Carbon ahead of competing models across all three cybersecurity benchmarks.
The announcement comes just two weeks after the launch of Gemini 4 Argon, Google’s previous frontier model. Carbon builds on that release with improvements in coding, reasoning, cybersecurity, and autonomous task execution.
Unlike conventional AI assistants that respond to individual prompts, Carbon is designed to carry out extended workflows involving multiple steps, tools, and decisions.
Google said its internal teams have observed the model working for days on complex engineering problems, including identifying the causes of software bugs and developing patches across large codebases.
However, these extended operations were reported by Google and have not been independently verified.
Stronger security and coding
Carbon’s cybersecurity capabilities cover three major stages of software security: finding vulnerabilities, determining whether they pose a genuine risk, and developing fixes.
On SEC-bench Pro, the model achieved 88.2%, while its 54.8% ExploitGym score reflects its ability to validate security vulnerabilities through controlled exploitation tests.
Its 71% CWE-bench result measures performance in generating working patches that address underlying software weaknesses.
Google also reported that Carbon more than doubled Argon’s success rate in generating secure code during internal testing.
The Chrome team found that Carbon performed more than 10 times better than a competing frontier model in producing high-quality security patches, according to Google. The company did not identify the competing model.
Beyond cybersecurity, Carbon ranked first on version 5.0 of the Artificial Analysis Intelligence Index, scoring 59.
It also improved its WebDev Arena score by 89 Elo points, reaching 1,768 and placing fourth on the leaderboard.
Google said the improvements extend to scientific tasks, multimodal understanding, and full-stack web development.
Safety improvements and rollout
Google has also introduced additional safeguards intended to reduce unreliable behavior during extended AI workflows.
In internal testing, Carbon reduced unsupported claims by approximately half compared with Argon. High-severity misalignment alerts fell by nearly 80% in live coding environments involving Google’s teams.
The company also reported a 0% attack success rate on Gray Swan’s Indirect Prompt Injection benchmark, compared with 0.7% for Argon.
These results reflect performance under specific testing conditions and do not establish that the model is immune to manipulation.
Google is initially restricting Carbon to a selected group of cybersecurity defenders, with additional trusted partners gaining access through its Fairwind Program.
The company is also participating in the US government’s voluntary pre-release model testing process.
Broader availability for developers, businesses, and consumers is planned, although Google has not announced a specific release date.
Carbon will carry the same introductory pricing as Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens will receive a 95% discount.
The phased rollout reflects the dual-use nature of advanced cybersecurity AI, which can help defenders identify weaknesses but may also provide capabilities that could be misused.