A flaw in the Linux kernel's TCP code, tracked as CVE-2026-98432, carries a CVSS score of 9.8 out of 10 and affects Ubuntu, Red Hat Enterprise Linux, and other distributions. The kernel's own CVE team assigned the score. Anthropic said on October 27 that its AI model wrote an exploit that uses the flaw to let a remote attacker with no credentials run commands as root on Linux servers with a specific configuration.
The vulnerability sits in the kernel's TCP receive path, the code that handles incoming network packets. An attacker who can reach a listening TCP port can send a crafted sequence of packets that corrupts kernel memory and ultimately yields root-level command execution. No authentication is required. The attacker does not need a local account, a stolen credential, or any prior access to the target machine.
The flaw matters now because of who can exploit it and how fast. Anthropic disclosed that its AI model generated a working exploit in roughly the time it takes to compile the kernel. That compresses a task that used to take weeks of specialized reverse engineering into an afternoon. Security teams at organizations running Ubuntu or RHEL now face a shorter window between disclosure and exploitation than they did a year ago.
What the flaw actually does
The bug is a memory corruption issue in the TCP receive path. When a server accepts a TCP connection, the kernel allocates and manages buffers to hold incoming data. Under specific conditions involving out-of-order packet delivery and a particular buffer size, an attacker can write past the end of a kernel buffer. That write lands in adjacent kernel memory, which an attacker can shape to control execution flow.
The result is remote code execution with root privileges. On a default-configured server the flaw is not reachable, which is the one piece of good news. The vulnerable code path only activates when a specific TCP option is enabled, a setting that some distributions turn on by default for performance reasons. Administrators need to check whether their systems have it enabled.
Red Hat published an advisory confirming the flaw affects RHEL 9 and RHEL 10. Ubuntu confirmed impact on versions 24.04 LTS and 26.04 LTS. Other distributions that ship the affected kernel code, including Debian and SUSE, are evaluating their exposure. Kernel maintainers merged a fix into the mainline tree on October 25, two days before Anthropic's disclosure.
Why the AI exploit changes the math
Security researchers have used AI models to find bugs for over a year. Writing a working exploit from a bug report is a different task. It requires understanding kernel memory layout, building a reliable trigger, and bypassing mitigations like KASLR and stack canaries. That work used to separate a disclosed vulnerability from a weaponized one by weeks or months.
Anthropic's model collapsed that gap. The company said the model produced a functional exploit that achieves root on a test system with the vulnerable configuration enabled. Anthropic did not release the exploit code. It shared the result with kernel maintainers and with CISA.
The practical effect is that defenders cannot rely on the old assumption that a disclosed CVE buys them time. When an AI can turn a patch commit into a working exploit in hours, the patch-to-exploitation window shrinks to however long it takes to download and install updates. Organizations that patch monthly are now exposed for weeks.
What administrators should do now
The fix is a kernel update. Both Red Hat and Ubuntu have released patched kernels. Apply them through your normal update channel, then reboot. Kernel updates do not take effect until the running kernel is replaced, so a reboot is mandatory.
If you cannot reboot immediately, disable the vulnerable TCP option as a mitigation. Red Hat's advisory includes the sysctl command to turn it off. Ubuntu's advisory does the same. The mitigation breaks no functionality for most workloads, though it may slightly reduce throughput on high-latency connections.
Check your exposure with a simple test. Run sysctl net.ipv4.tcp_recovery and sysctl net.ipv4.tcp_fastopen. If either returns 1, your system has the vulnerable code path enabled. Set both to 0 as an immediate stopgap, then schedule the kernel update and reboot.
The broader pattern
This flaw is the third high-severity Linux kernel vulnerability disclosed in 2026 that an AI model helped exploit. The first, in March, was a use-after-free in the networking stack. The second, in July, was a race condition in the virtual file system layer. Each one was patched within 48 hours of disclosure, but each one also saw active exploitation in the wild within a week.
The pattern is not that Linux is becoming less secure. The kernel's code quality has improved steadily, and the number of CVEs per release has dropped. The pattern is that the cost of turning a disclosed bug into an attack has fallen to near zero. A single researcher with an AI subscription can now do work that used to require a team.
That shifts the burden onto defenders. Patching faster is necessary but not sufficient. Organizations need to reduce their attack surface by disabling features they do not use, segmenting networks so that a compromised server cannot reach others, and monitoring for the specific packet patterns that trigger this flaw. CISA published detection guidance on October 28.
What comes next
The kernel maintainers are reviewing whether to make the vulnerable TCP option opt-in rather than opt-out in future releases. That change would reduce the number of systems exposed by default, though it would not help the millions of servers already running with the option enabled.
Anthropic said it is working with the kernel security team to develop a process for AI-assisted vulnerability research that includes coordinated disclosure. The goal is to give defenders a head start before exploit code circulates. Whether that process works in practice depends on how fast AI capabilities improve between now and the next disclosure.
For now, the advice is simple. Patch, reboot, and check your sysctl settings. The flaw is real, the exploit exists, and the window is shorter than it used to be.
For the full technical analysis, see the kernel security advisory.
Related coverage on this site: cybersecurity desk.