There's a moment in security research when a new class of tool arrives and everyone quietly updates their threat model. We may be at one of those moments.

Google this week unveiled Gemini 4 Carbon, its latest frontier model, and the headline numbers are hard to ignore. It scored 88.2% on SEC-bench Pro — the industry's standard test for finding vulnerabilities — along with 54.8% on ExploitGym, a benchmark for actually weaponising those findings. No model has led all three major cybersecurity benchmarks simultaneously before. Google is rolling it out first to vetted defenders through something it's calling the Fairwind Program, which is the right instinct even if "vetted defenders" is doing a lot of work in that sentence.

Then Anthropic landed its own headline. An unreleased Claude model — not the one you can sign up for today — has been demonstrated finding and chaining multiple vulnerabilities across Chrome, Safari, Android, Windows, and Linux. Separately, that same model reportedly wrote a functional exploit for the critical Linux kernel TCP flaw (CVE-2026-98432, CVSS 9.8) that was disclosed this week, giving a remote, unauthenticated attacker root on affected servers.

To be clear: Anthropic disclosed both findings itself. That's worth noting. The company is leaning into transparency about what its models can do offensively, presumably because it believes the alternative — pretending capability doesn't exist — is more dangerous. It's a reasonable bet, though not a cost-free one.

What this actually means for defenders is less dramatic than it first reads, but still important. AI doesn't change the physics of patching — you still need to find the vulnerable system, push the update, verify it landed. What it changes is the economics of finding vulnerabilities in the first place. Offensive researchers and nation-state actors with access to capable models can now compress months of manual vulnerability research into hours. The gap between "patch released" and "exploit in the wild" was already shrinking. These benchmarks suggest it could narrow further.

The Linux kernel TCP flaw is a good test case. CVE-2026-98432 affects Ubuntu, Red Hat Enterprise Linux, and other major distributions — the kind of infrastructure running under a significant portion of the internet, including systems operated across Australian government and enterprise environments. A patch exists. The question, as always, is how quickly it gets deployed.

What to watch: Whether Google's Fairwind Program actually gates access effectively, or becomes another well-intentioned control that sophisticated actors route around. Also watch how quickly CVE-2026-98432 patches propagate — the combination of a 9.8 CVSS score and a demonstrated AI-generated exploit makes this one of the more urgent patch cycles of the year.

The broader picture is one the security industry has been nervously anticipating: AI models that can both find vulnerabilities at scale and generate working exploit code are no longer theoretical. They shipped this week. The defenders using the same tools have a narrow window to close the gap before the attackers do.