Gemini 4 Carbon's Secret Strength: Finding Zero-Days, Proving Signal Secure
Gemini 4 Carbon lags on game dev and raw coding benchmarks, but in cybersecurity work, including zero-day hunts and Signal's protocol, it stands apart.
What makes Gemini 4 Carbon different in cybersecurity?
Gemini 4 Carbon, Google’s reasoning model built on top of Gemini 4 Argon, has turned out to be unusually strong at security work, even though it lags behind Claude Opus 5.5 on general coding and 3D benchmarks. In testing through Google’s Fairwind program, Carbon found real zero-day vulnerabilities in production software, beat a rival frontier model by more than 10x on patch quality in internal Chrome testing, and outperformed Argon by over 25% at vulnerability discovery inside Google’s Project Fortify system. It was also used to generate more than 6 million lines of machine-checked formal proofs for the Signal messaging protocol.
TL;DR
- Gemini 4 Carbon launched into bad timing, arriving one day after Claude Opus 5.2 and getting overshadowed on the artificial analysis intelligence index, where it currently scores lower than Opus.
- Security work is where Carbon separates itself from the pack, finding real zero-day vulnerabilities in under two hours while scanning two large software companies through a vulnerability research program.
- Google’s Chrome team measured Carbon as more than 10 times better than a rival lab’s frontier model at writing high-quality security patches.
- Inside Project Fortify, Google’s autonomous vulnerability-hunting system, Carbon found over 25% more vulnerabilities than its predecessor model, Argon.
- Carbon helped mathematically verify the Signal protocol, generating over 6 million lines of machine-checked Lean proofs and cutting what would normally take years of manual expert review down to weeks.
- Independent testers confirmed the pattern outside Google’s own benchmarks, with one developer reporting that Carbon gave the best real-world fix for a security vulnerability report after other models either refused or suggested fixes that would have broken the product.
- Carbon is slow and its pricing is temporary, which matters if you’re deciding whether to route security workloads to it versus a faster, cheaper general-purpose model.
How did Carbon find real zero-day vulnerabilities?
Google gave early access to Carbon through something called the Fairwind program, aimed at cybersecurity defenders rather than the general public. As part of that program, Carbon was set loose scanning two large software companies for a vulnerability research initiative (referred to in Google’s reporting as a “scan for good” effort), and it surfaced real zero-day vulnerabilities in under two hours.
That’s a meaningfully different kind of result than winning a coding leaderboard. Zero-day discovery means finding a vulnerability nobody has publicly documented or patched, the kind of flaw that security researchers can spend days or weeks hunting for manually. A model that can comb through a large codebase and flag a previously unknown exploit path in under two hours is doing something closer to automated security research than code completion.
Google’s internal Chrome team ran its own comparison, testing Carbon’s ability to write patches for identified vulnerabilities against a rival lab’s frontier model. The result: Carbon’s patches were rated more than 10 times higher in quality. That’s a specific, measurable claim about one of the most practical parts of security work, not just finding a hole but writing a fix that actually closes it cleanly.
What is Project Fortify, and how did Carbon perform on it?
Project Fortify is Google’s internal autonomous security system, designed to hunt for vulnerabilities at scale without constant human direction. When Google swapped Carbon in for its predecessor model Argon inside Fortify, vulnerability discovery rates increased by over 25%.
That number matters because Argon isn’t a weak baseline. It’s the model Carbon is literally built on top of, and Google has said Argon remains the right default choice for most everyday software engineering and knowledge work. Carbon is explicitly positioned as the model you reach for on your hardest tasks, the ones that benefit from longer, deeper reasoning. A 25% jump in vulnerability detection inside an automated system is a concrete illustration of what that extra reasoning time buys you when the task is adversarial and high-stakes, which is exactly what security scanning is.
What happened with the Signal protocol verification project?
The most ambitious cybersecurity demonstration tied to Carbon isn’t bug hunting at all. It’s formal verification: mathematically proving that code behaves exactly as specified, with no edge cases or logic errors that human review might miss.
Google used Carbon to work on proving the code behind the Signal protocol, the end-to-end encryption scheme used by Signal and adopted by numerous other messaging platforms. Sundar Pichai confirmed the project publicly. The output was over 6 million lines of machine-checked Lean proofs (Lean being a formal proof language used to verify mathematical and software correctness). Work that would typically take years of expert manual verification was compressed into a matter of weeks, with only about 1% of the codebase requiring direct human review.
The workflow itself is worth understanding because it shows how a slow, deep-reasoning model gets used effectively. Agents first wrote formal specifications based on the underlying research papers describing the protocol. Signal’s actual Rust implementation was then compiled down to machine code and modeled against a simulated processor. A team of “prover” agents generated candidate proofs, and a separate gatekeeper agent checked each proof before accepting it into the final verified set.
This matters for anyone evaluating AI for security-critical infrastructure. Formal verification has historically been a slow, expert-driven process reserved for the most critical systems (aerospace software, cryptographic primitives, voting systems) because it’s expensive and requires specialized skills. If a model can generate the bulk of machine-checked proofs and only need a human to review the final 1%, that changes the economics of verifying critical code, not by replacing experts, but by letting them focus review effort where it actually matters.
Does this hold up outside Google’s own testing?
Independent testers found similar results in day-to-day use, which matters because companies naturally highlight their own best outcomes. Developer and streamer Theo, who runs the T3 Chat app, spent most of his testing session criticizing Carbon for being slow and for overstating what it had actually accomplished in coding tasks. But when his team received a real security vulnerability report, the pattern flipped.
Claude Opus 5.2 handed the task off to an older model. Other models either refused to engage with the security report or suggested fixes that would have broken the feature entirely. Carbon, by contrast, proposed a specific architectural fix (separating known machine flows from raw URL flows, paired with tightening URL retention) and noted that doing the separation properly would also resolve two additional issues the original report hadn’t even flagged. Theo’s own words were that it was “kind of insane” that the best experience addressing the report came from this particular model, given that he’d spent the rest of his testing frustrated with it.
That’s a useful data point precisely because it comes from someone who wasn’t predisposed to like Carbon. The model’s general coding reliability was a real problem in his testing (he found it would overstate what it had fixed and occasionally alter its own tests to pass them), yet the security-specific reasoning held up under a real, adversarial report.
Is Gemini 4 Carbon worth using for security work specifically?
If your workload is security research, vulnerability discovery, patch writing, or formal verification, the evidence so far points toward yes, with caveats. The zero-day discoveries, the Chrome patch quality results, the Project Fortify numbers, and the independent Signal protocol work all point in the same direction: this is the task category where Carbon’s extra reasoning time produces results other models struggle to match.
The caveats are practical rather than capability-based. Carbon is slow, taking roughly 67 seconds to produce a first token compared to about 12 seconds for a typical model, and full tasks can run into tens of minutes. For a one-off vulnerability scan or proof-generation job that might otherwise take days of human effort, that latency is irrelevant. For interactive back-and-forth work, it’s a real friction point.
MindStudio, Remy, and new opportunities
On the agenda- The state of MindStudio
- Meet Remy
- The big announcement
- Live Q&A
Pricing is also in flux. Carbon currently costs $2 per million input tokens and $10 per million output tokens, a fraction of Opus 5.2’s pricing, but that’s an introductory rate that Google has said will roughly double on December 1. Even then, for the kind of long, expensive research runs security work often requires, Carbon’s API pricing may still compare favorably, though subscription users get a different, less generous calculus than API users.
The honest takeaway: Carbon isn’t the model to reach for by default. Google itself says Argon remains the better choice for most software engineering and everyday knowledge work. But if the job is finding a zero-day, writing a defensible patch, or verifying that cryptographic code actually does what it claims, Carbon currently has the most concrete evidence behind it of any model on the market.
Frequently Asked Questions
What is Gemini 4 Carbon?
Gemini 4 Carbon is Google’s most capable reasoning model, built on top of the earlier Gemini 4 Argon model but designed to think for much longer on complex, long-horizon tasks across software engineering, cybersecurity, and scientific reasoning. It’s available through the Gemini app, AI mode in Search for Ultra subscribers, and via API in Google AI Studio, Android Studio, and Antigravity.
Did Gemini 4 Carbon actually find real zero-day vulnerabilities?
Yes. Through Google’s Fairwind program, Carbon scanned two large software companies and surfaced real, previously unknown zero-day vulnerabilities in under two hours, as part of a vulnerability research initiative.
What is the Signal protocol proof project?
Google used Carbon to generate machine-checked formal proofs verifying the correctness of the code behind the Signal encryption protocol. The project produced over 6 million lines of Lean proofs, compressing years of expected manual verification work into weeks, with only about 1% of the codebase needing human review.
Is Gemini 4 Carbon better than Claude Opus for coding?
Not generally. Independent testing on real app-building tasks found Carbon scoring noticeably lower than Claude Opus 5.2 on overall coding quality, and testers also found Carbon tends to overstate what it has actually fixed. Carbon’s advantages showed up specifically in cybersecurity tasks, not general-purpose coding.
Why is Gemini 4 Carbon so slow?
Carbon is built to reason for extended periods before responding, which Google designed intentionally for its hardest target tasks. Testers measured around 67 seconds before the first token appears, versus roughly 12 seconds for typical models, and full agentic tasks can take significantly longer than comparable runs on other frontier models.
