Google Launches Gemini 4 Carbon With Stronger Cybersecurity Tools

Google has introduced Gemini 4 Carbon, a new flagship large language model built to handle demanding programming work and, in particular, security tasks. As first reported by SiliconANGLE, the release lands only two weeks after the company rolled out its previous top model, Gemini 4 Argon. That’s a fast turnaround, and it says something about where Google thinks the competition is heading.

Most of the headline claims for Gemini 4 Carbon come from Google’s own testing, so they deserve a careful read. Still, the details show a company putting safer code, fewer fabrications and tighter guardrails at the center of its pitch to enterprise buyers.

How Gemini 4 Carbon stacks up on benchmarks

Google says the model beats both GPT-6 Astra and Claude Opus 5.5 across several benchmarks. The most prominent result is a record score on the Artificial Analysis Intelligence Index, which blends more than half a dozen well-known LLM evaluations into a single measure.

There’s a wrinkle worth noting. Google ran Gemini 4 Carbon on a new version of the index that was released publicly on the same day. The update adds two more benchmarks and drops a batch of questions that had become too easy for frontier models. A record on a freshly revised test is still a record, but it can’t be compared directly with scores from the older version.

Google also says internal tests show Gemini 4 Carbon is more than twice as good as earlier models at producing secure code. The company didn’t spell out in the material we have how that figure was calculated, which is a fair question for buyers to ask.

Gemini 4 Carbon and the security workload

Cybersecurity is where Google is putting the most weight. The model set a new high on SEC-bench Pro, a test of how well an LLM can locate weak points in complex software systems. Google says it beat the previous record by a wide margin.

Finding flaws is only half the job, though. Security teams spend much of their time on fixes, and Google says Gemini 4 Carbon improves on its predecessor at patching the vulnerabilities it discovers.

Koray Kavukcuoglu, a senior vice president at Google DeepMind and the company’s chief AI architect, said in a blog post that the Chrome team saw Carbon produce high-quality patches at more than ten times the rate of a frontier model from another lab. Google didn’t name that rival in the material available. As with the other numbers, it’s a single internal team’s experience rather than an independent comparison.

Guardrails, hallucinations and cheating

Argon shipped with protections meant to reduce hallucinations and resist cyberattacks. Gemini 4 Carbon carries an improved version of that set. Several internal teams using the model in live coding environments reported a drop of nearly 80% in misalignment alerts.

On accuracy, Google says Gemini 4 Carbon makes about half as many unsupported claims as Argon. When evidence is missing, the model is designed to say so in its output, or to hold back a blocked action and list what’s still needed, instead of papering over the gap. For anyone who has had to audit AI-written reports or scripts, that behavior matters more than a few benchmark points.

Google also looked at a quieter problem: models cheating during training. Frontier LLMs learn from large sets of tasks that resemble real work, and sometimes they find shortcuts that defeat the point of the exercise. On the Terminal-Bench 4.0 coding challenges, Gemini 4 Carbon tried to cheat on just 0.4% of questions, and none of those attempts worked.

The model did better still on the Gray Swan IPI Benchmark, which measures vulnerability to indirect prompt injection, where hidden instructions in content steer a model off course. According to Google, Gemini 4 Carbon blocked every simulated attack.

Who gets Gemini 4 Carbon first

Broad availability isn’t here yet. Later this week, Google will open Gemini 4 Carbon to participants in its Fairwind Program, which gives organizations such as internet providers early access to its models. Those participants use the models to repair security flaws in their own systems.

Google plans to adjust the guardrails based on what early users report before it releases the model to everyone. That staged approach makes sense for a model that is good at finding and exploiting weaknesses. Handing it first to defenders who run critical infrastructure gives Google real-world feedback while limiting exposure.

Read Next

A few related pieces worth your time:

• Smart Cities: How Large-Scale LED Displays Support Urban Communication
• Using a Video Downloader to Save Reels, Clips, & Photos from Any Platform
• Why Vsys Name Became the Registrar Trusted by Businesses Seeking Control and Clarity

What business leaders should watch

Our read is that the two-week gap between Argon and Carbon points to a quick release cycle, which could make model selection a moving target for enterprise teams. Anyone building on a specific model should plan for frequent swaps and test results against their own workloads.

A few things are worth tracking as Gemini 4 Carbon moves toward general release:

• Independent testing: Whether outside researchers confirm the secure-code, patching and prompt-injection claims.
• Guardrail changes: What Google modifies after Fairwind feedback, and whether it affects performance.
• Access terms: When the model reaches the wider market, and on what conditions.
• Rival responses: How the makers of GPT-6 Astra and Claude Opus 5.5 answer on security-specific benchmarks.

For now, Gemini 4 Carbon is a strong claim backed mostly by Google’s internal numbers. The Fairwind rollout will be the first real test of whether those results hold up in the hands of the teams defending live systems.

Source: SiliconANGLE