Google debuts flagship Gemini 4 Carbon model with cybersecurity upgrades - SiliconANGLE
Google LLC has debuted Gemini 4 Carbon, a frontier large language model that excels at complex programming tasks and cybersecurity applications.
The LLM is rolling out just two weeks after the search giant introduced its previous flagship model, Gemini 4 Argon. According to Google, internal tests have shown Carbon to be more than twice as good at generating secure code. Additionally, the model outperforms GPT-6 Astra and Claude Opus 5.5 across multiple benchmarks.
Benchmark Performance
Carbon achieved a record score on the Artificial Analysis Intelligence Index, a test combining more than a half-dozen popular LLM evaluations. Google tested Carbon on a new version of the index that was publicly released today, which includes two additional benchmarks and removes saturated questions deemed too easy for frontier models.
Cybersecurity Focus
The model is particularly adept at cybersecurity tasks, handily beating the previous record on SEC-bench Pro, which evaluates LLMs' ability to find weak points in complex software systems. Carbon is also better than its predecessor at patching the vulnerabilities it finds.
The Chrome team has reported that Carbon performed more than 10X better than another lab's large frontier model when generating high-quality patches
Improved Guardrails
Google implemented improved guardrails in Carbon to mitigate hallucinations and cyberattacks. Internal teams using Carbon in "live coding environments" reported a nearly 80% drop in misalignment alerts. The model is configured to avoid generating inaccurate information when it lacks knowledge.
In internal testing, Carbon cuts unsupported claims by approximately 2x over Argon. It puts methodological caveats directly into deliverables or withholds blocked actions with a clear list of requirements, rather than fabricating information around missing evidence.
Training Integrity
Google tested whether Carbon would use shortcuts to solve coding challenges in the Terminal-Bench 4.0 benchmark. The model attempted to cheat across only 0.4% of questions, with none successful. Carbon performed even better on the Gray Swan IPI Benchmark, blocking all simulated prompt injections.
Availability
The company will make Carbon available later this week through its Fairwind Program, which gives organizations such as internet providers early access to its LLMs. Participants use the models to fix cybersecurity flaws in their systems. Google plans to refine Carbon's guardrails based on feedback from early adopters before making it broadly available.
Related
