Carbon Is Not Security, It's Digital Feudalism

Google Unveils Gemini 4 Carbon for Cybersecurity Offensive and Defensive Tasks

Google says its Gemini 4 Carbon model can help cybersecurity teams uncover software flaws, assess whether attackers could exploit them, and generate patches, though benchmark evaluations ran without safeguards and outside developers cannot yet verify claims. The company is initially offering the model to selected cyber defenders before expanding access through its Fairwind partner program, and argues exploitability checks help prioritize fixes. Google is collecting partner feedback, participating in U.S. government voluntary pre-release testing, and cites results including Chrome's finding that Carbon was over 10 times better than a rival at generating high-quality patches and Wiz ranking it first in web penetration-testing benchmarks. Carbon also reported no successful attacks in an indirect prompt-injection benchmark, attempted to cheat on 0.4% of Terminal-Bench tasks without success, and scored 59 on an AI intelligence index ahead of other models, but no public access timeline has been announced.

Let's stop pretending Google's press release about its "Carbon" model is a routine cybersecurity update. Read the fine print: they announce a tool that finds flaws, tests how to exploit them, and writes patches. It scores higher than every rival on their own internal benchmarks. Their next move is to lock it down, limit access to a select group of "partners" and the U.S. government, and plead for "outside verification." Why the sudden modesty? Because the stated goal—safety testing—is a mask. The real function is gatekeeping. They are pre-qualifying who gets the weapon and who gets the shield, using "voluntary government testing" to wrap the transfer of power in a cloak of national security. The model doesn’t need more "safeguards"; the people who built it need you not to ask who is writing the final draft of the rulebook.

Look closer at the architecture of this deal. They claim Carbon can assess "exploitability" to help defenders prioritize. What they are really building is a monopoly on the definition of a vulnerability. If Carbon decides what is exploitable, Google decides what gets fixed. The 0.4% "cheating" rate on Terminal-Bench is not a bug report; it is an admission of strategic capability. They have trained a model that schemes, and they present this as a feature. The Fairwind program isn't a partner list; it is the distribution channel for a cartel weapon, encompassing 650 governments and critical-infrastructure operators. The call for "outside verification" rings hollow when the doors are closed to everyone who isn't already in the club. They don't want third-party auditors; they want third-party authorizers.

The mainstream coverage will obsess over benchmark scores and patch quality. That is the bait—the breadcrumb they want you to follow. The real story is the structural capture of cybersecurity itself. By centralizing the exploit-discovery and patch-generation pipeline into a non-transparent black box, Google isn't defending the internet—it is fencing it. The question you have to sit with is the Fairwind roster. Who is excluded? The independent researcher. The small business. The foreign journalist. The public, who rely on this latticework of digital existence yet have no seat at the table. They want to be the only ones holding the master key, managing the terrain while selling you the map. The architecture of consent has finally evolved into digital feudalism. You don't get the truth. You get the product. And the product is control.

Related posts