x
Cybersecurity has reached an unprecedented turning point. Artificial intelligence is no longer merely finding minor software glitches; it is crafting sophisticated, fully functional exploits capable of taking over enterprise servers without password authentication. As highlighted in a breaking report by Decrypt, Anthropic revealed that an internal research version of its Claude AI successfully developed a zero-click remote code execution attack against the Linux kernel.
The flaw, tracked officially as CVE-2026-98432, carries a maximum severity CVSS score of 9.8 (Critical) and has been quietly lurking in Linux kernel releases since version 4.4. While major enterprise distributions like Red Hat moved swiftly to issue patches over the weekend, millions of systems running Ubuntu remain exposed with no official fix yet deployed.
In a detailed breakdown published in Anthropic's research post, researchers demonstrated how Claude discovered and exploited the flaw with minimal human prompting. The AI managed to orchestrate a complex two-stage attack chain against servers exposing a Multipath TCP (MPTCP) listener.
First, Claude bypassed Kernel Address Space Layout Randomisation (KASLR) memory protections. In just 48 minutes, Claude exploited a subtle race condition in the kernel's IPv4 networking stack. By forcing the kernel to leak memory addresses, the AI accurately mapped out the kernel's memory layout.
Five minutes later, Claude leveraged a use-after-free (UAF) flaw to overwrite memory callback pointers, executing arbitrary shell commands with complete root privileges.
What makes this vulnerability particularly alarming is its forgiving nature for attackers. If a race condition fails, the connection handshake simply drops without crashing the host server. This allows an automated script to retry the attack indefinitely until it succeeds, making remote compromise virtually inevitable over time.
To understand why this flaw is so dangerous, it helps to examine how the Linux kernel handles network traffic. As explained by AlmaLinux OS, when an incoming TCP connection request (a SYN packet) arrives containing IPv4 options, the kernel temporarily stores those options to echo them back in its SYN-ACK response. When the socket handshake completes, ownership of these options is transferred to the child socket without making a fresh copy.
If a retransmitted SYN packet or socket migration occurs simultaneously on another CPU core, the socket can be freed using standard kfree() memory routines rather than waiting for an RCU grace period (kfree_rcu()). This leads to a classic use-after-free race condition inside ip_options_build(), allowing an attacker to manipulate freed memory blocks.
While Anthropic's demonstrated attack relied on Multipath TCP, security experts warn that system administrators should not assume they are safe just because MPTCP is turned off.
In an official advisory published on the oss-security mailing list, the Go Security Team pointed out that while Go 1.24 introduced MPTCP enabled by default for Linux listening sockets, the core vulnerability resides in general IPv4 TCP connection handling. Disabling MPTCP in application settings or Go build flags provides only a fragile workaround. The flaw can potentially be triggered on standard IPv4 TCP listeners, making a full kernel update the only definitive cure.
The disclosure of CVE-2026-98432 triggered a flurry of activity across major Linux enterprise vendors, revealing stark differences in response times across the industry.
Red Hat acted quickly on Sunday, issuing security advisory RHSA-2026:93412 to patch Red Hat Enterprise Linux (RHEL) versions 8, 9, and 10.
AlmaLinux OS technical steering committees authorised backporting the upstream kernel fix by Linux developer Eric Dumazet immediately, releasing production kernel updates across AlmaLinux 8, 9, and 10 ahead of upstream RHEL builds.
In stark contrast, Ubuntu's security tracker still lists virtually every release—from Ubuntu 16.04 LTS up to the latest 26.10 release—as vulnerable. Fixes for active LTS releases (such as 22.04 and 24.04) remain listed as work in progress, leaving millions of cloud instances and servers exposed to potential exploitation.
If you are managing Ubuntu servers or systems where immediate kernel reboots are not feasible, temporary packet filtering is critical. As outlined in Red Hat's vulnerability bulletin RHSB-2026-016, dropping incoming IPv4 TCP packets that carry IP options stops the exploit trigger before it reaches the vulnerable kernel stack.
Administrators can set up an nftables input rule to drop any IPv4 TCP packets where the IP header length exceeds 5 words. This filters out the malicious payload without disrupting regular TCP traffic, which relies on TCP header options rather than IPv4 options. However, this is strictly a temporary stopgap, and system operators must prioritise upgrading their Linux kernels to patched builds as soon as official vendor updates arrive.
Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.
