Anthropic says an unreleased Claude model found a Linux kernel vulnerability that lay dormant for 11 years
Anthropic published a security research report stating that its internal unreleased Claude model, during systematic testing, discovered a critical Linux kernel vulnerability that had lain dormant for 11 years, CVE-2026-98432, with a CVSS score of 9.8, at the remote code execution level. The vulnerability stems from a use-after-free flaw in the IPv4 TCP handshake; the code was submitted in October 2015 and introduced with Linux 4.4, and an attacker can remotely obtain Root privileges without a password or user interaction. Claude also generated a working two-stage exploit: the first stage took about 48 minutes to defeat kernel address space randomization, and the second stage took 5 minutes to execute arbitrary commands with Root privileges, for a total of about 53 minutes. The upstream fix was completed on October 15, but as of November 1, Ubuntu 22.04 LTS, 24.04 LTS, 26.04 LTS, and 26.10 are still marked as affected.
Why read this
AI score 77
Anthropic disclosed that an unreleased Claude model found a Linux kernel vulnerability that lay dormant for 11 years and generated a working exploit, illustrating the boundaries of AI capabilities in vulnerability discovery and attack chains.
Topics
Body · AI translation
Add crowdsourcing promotion copy Add crowdsourcing promotion copy end Add App-exclusive traffic redirect
November 2 news: Anthropic recently released security research disclosing that its internal, unreleased Claude model, during systematic testing, discovered a severe vulnerability that had been lurking in the Linux kernel for 11 years.
The vulnerability is numbered CVE-2026-98432, with a CVSS severity score as high as 9.8, belonging to the highest-risk category of remote code execution.
The root cause of the vulnerability is a use-after-free (UAF) flaw in the IPv4 TCP handshake. The code was submitted in October 2015 and introduced with the Linux 4.4 kernel; when handling TCP connection retransmissions, one CPU may read memory that another CPU has already freed.
An attacker needs no password and no user interaction, and can trigger it solely through network traffic, thereby achieving remote Root privilege execution. The kernel CVE team characterized it as a severe vulnerability that is network-reachable, requires no privileges, and requires no interaction.
Even more alarming is the role of AI in this. Anthropic's Claude model not only discovered the vulnerability but also generated a working two-stage exploit. The first stage took about 48 minutes, defeating kernel address space randomization protection; the second stage took only 5 minutes to execute arbitrary commands with Root privileges. The entire attack process takes about 53 minutes, and a failed exploit does not crash the target, so attackers can try repeatedly.
The upstream fix was completed on October 15, covering versions 5.10.272, 6.1.190, 6.12.113, 6.18.56, 7.2.10, and 7.3-rc7. Distributions such as Debian, Red Hat, and Amazon Linux have successively released patches.
But the situation with Ubuntu is completely different: as of November 1, Ubuntu 22.04 LTS, 24.04 LTS, 26.04 LTS, and 26.10 are still marked as affected, and fixes are still under development.
For Ubuntu administrators running externally facing services, the risk window remains open. Go 1.24 and later versions enable Multipath TCP by default on Linux, which may expand the exposure surface of this vulnerability, but disabling MPTCP can only reduce risk and cannot eliminate the kernel-level flaw.
Anthropic stated that the exploit code has not been publicly released, but this incident has clearly demonstrated a trend: AI is no longer just a tool for writing code; it can also become a weapon for breaking down system defenses. [Source: Kuai Technology] Statement: This is an exclusive article from Sina.com. Reproduction without authorization is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin
Non-targeted 300*250 button end