Researcher Says AI Helped Develop Linux Traffic Control Race Into Root Exploit
STAR Labs has published a Linux kernel exploit that turns a standard local user into root on a CentOS Stream 9 target build. Error, followed by CVE-2026-53264 (CVSS score: 7.8), a post-implementation race in the kernel’s network traffic management subsystem.
Researcher Lee Jia Jie said that artificial intelligence (AI) helped him find the bug and speed up the development of the exploit. This is elevation of privilege, not remote code execution, so an attacker needs to hold a location on the machine before any of it can work.
The exploit shown also requires disallowed user namespaces, i CONFIG_NET_ACT_GACT again CONFIG_NET_CLS_FLOWER kernel options, and a kernel-specific return-oriented programming (ROP) chain that contains hard-coded offsets. Those scenarios quickly mitigated exposure, but the source code for the full exploit is now public.
The upstream fix arrived on June 1, 2026, and has since been backported to several stable kernel branches. The Linux CNA registry lists vulnerable versions starting with Linux 4.14. The fixed releases are 5.10.259, 5.15.210, 6.1.176, 6.6.143, 6.12.94, 6.18.36, and 7.0.13, where a major line fix goes into 7.1-rc7.
Linux users should install the kernel distribution that carries the fix rather than relying on the upstream version number alone. Hacker News found no accidental entry in CISA’s Known Exploited Vulnerabilities catalog and no official report of an exploit in the wild as of 28 July 2026.
Lee said in a technical post that AI helped with vulnerability discovery, the production of a Kernel Address Sanitizer (KASAN) proof of concept, and optimization of the race window. STAR Labs has also released exploit code aimed at CentOS.
Without a model, data, service, or interaction record, the disclosure is difficult to use as a benchmark for AI capabilities or to separate the system’s contribution from Lee’s guidance and judgment.
Hacker News has asked STAR Labs for details on the AI system, testing environment, and disclosure timeline and will update this story with any response.
“AI still has a lot of blind spots and failings in reasoning,” Lee said, adding that human judgment is still necessary for every task.
The vulnerability lies in the lifecycle management of Linux traffic control actions. Once RTM_NEWTFILTER again RTM_DELTFILTER functions can leave one thread reading an action object after another has released it. The upstream patch fixes the race by postponing free execution until the current read-copy-review (RCU) has finished.
The exploit creates its own namespaces for the user and network, providing a namespace CAP_NET_ADMIN without requiring host administrator privileges. It accesses the vulnerable path by using a clsact qdisc again flower filter. The Timerfd and epoll functions extend the race window, while the priority payment key assignment re-claims the freed object. The ROP chain then overwrites core_pattern.
The exploit places a copy of itself in memfd and crashes the child process on purpose, causing Linux to use the binary supported by memfd as a default dump handler in the original namespace.
Lee reported that the exploit succeeded in all 10 of his tests, taking between nine and 111 seconds on a laptop running CentOS Stream 9. Those reliability figures have never been independently reproduced. The fixed exploit gadget offset also means that it must be rebuilt for some kernel packages and may not be able to adapt to some new architectures.
The actual risk is less than the usual “Linux root exploit” label might suggest, but the public exploit code raises the urgency of parallel systems that remain unpublished.
The ascending episode introduces Kyle Zeng, who goes by the handle KyleBot, as a reporter. Lee said he discovered the bug independently and later discovered that Zeng had reported it just before the TyphoonPwn 2026 competition. Lee published the latest technical analysis and exploit code.
The distribution situation remained uneven as of July 28: Debian lists stable kernels for supported stable releases, Ubuntu still marks many packages as vulnerable, and SUSE lists the issue as pending across multiple products. SUSE separately assigned the bug a score of 5.5, using a vector that records only the availability effect, under Linux CNA’s 7.8 test.
Those trackers show the status of the package, not how many deployed systems have the required namespaces, kernel options, and corresponding kernel builds. Public sources reviewed do not indicate the number of people at immediate risk.
Lee wrote that the AI-heavy process makes bug hunting “feel like I’m doing n-day analysis of even new bugs.”
