AI Found a Kernel Flaw in 53 Minutes—Was It Planted?

Anthropic's Claude Research Model Discovers Critical Linux Kernel Vulnerability

Anthropic revealed that an unreleased Claude research model identified and exploited CVE-2026-98432, a use-after-free vulnerability in the Linux kernel’s IPv4 TCP processing that allows unauthenticated remote root access to servers running a Multipath TCP (MPTCP) listener, with the flaw present since kernel version 4.4; the proof-of-concept achieved root command execution in about 53 minutes, though CloudLinux could not reproduce the claim, and while an upstream fix is available, distributions like Ubuntu and Red Hat Enterprise Linux remain unpatched as of late October—administrators must install their distribution's fixed kernel, as disabling MPTCP alone does not address the underlying TCP issue, and Go 1.24’s default MPTCP enablement may inadvertently expose services.

The Managed Disclosure of a Planted Backdoor

Anthropic announces that a Claude research model found a kernel memory corruption flaw in under an hour — a use-after-free bug lurking since kernel 4.4, buried in the IPv4 stack of every Linux server running Multipath TCP. Let me stop you right there. An AI model, trained on the sum of human code, zeroes in on a vulnerability that somehow escaped every human audit, every static analysis tool, every fuzzing campaign for years — and then demonstrates a working remote root exploit in 53 minutes? That is not discovery. That is retrieval. That code was placed there, waiting for the right moment, and the model was simply handed the map. The fact that CloudLinux's kernel team "could not reproduce the claim" is not a refutation — it's the tell. When the institution that would be implicated says the exploit doesn't work, you ask who benefits from that denial. The consensus machinery is already spinning.

The Coincidence That Isn't: Go 1.24 and the MPTCP Trojan Horse

Now look at what else the article buries in the second half: Go 1.24 enables MPTCP by default for TCP listeners on Linux. That means every server running a Go service — and that is a very large slice of the internet — suddenly exposes the exact attack surface this flaw exploits, without the administrator ever choosing to. You have to ask yourself: who pushed for that default? Who sat on the kernel steering committees? Who wrote the MPTCP specification? The same people who funded the AI lab, the same people who sit on the boards of the foundations that manage the Linux kernel's governance. This is not a bug. This is a perception shepherding exercise — a demonstration that they can discover and control the vulnerabilities they have seeded, while publicizing it as a heroic act of "research." The upstream fix is already out, but the patch delays for Ubuntu and Red Hat are not technical delays. They are timed delays, keeping systems open until the next phase of the operation.

Your Children, Your Servers, Your Future

The severity score is 9.8 — a "critical" that gives an unauthenticated remote attacker root on your server. But Red Hat calls exploitation "complex" and pretends denial of service is the real risk. That is the script. That is how they manage the narrative: downplay the danger so you don't panic, while the same elites who control the AI, the kernel, and the distribution channels hold the keys. The villain is not a lone hacker — it is the interlocking network of foundation philanthropists, AI safety theater, and open-source governance that treats your infrastructure as a laboratory. The question you should be sitting with tonight is not whether to patch. It is: who specifically pushed MPTCP into the default kernel configuration? Who chairs the technical advisory board of the Linux Foundation? Look up those names. Trace the grants. Follow the money from the same foundations that fund Anthropic. The answer is already in front of you — they told you themselves in the very paragraphs of this "news."

Related posts