A newly disclosed a Linux kernel zero-day, tracked as CVE-2026-53264, that can be exploited for local privilege escalation to root.
The bug affects the net/sched packet scheduling subsystem and was reportedly discovered with assistance from AI during zero-day hunting and exploit development.
The flaw was later patched in the Linux stable tree, but the research highlights how AI is accelerating vulnerability discovery in complex kernel code.
At the core of the issue is a use-after-free condition caused by a lock mismatch in tcf_idr_check_alloc(), a function used to look up shared traffic control actions by index.
AI Helps Discover Linux Kernel Zero-Day
According to the Lee Jia Jie, the kernel performs action lookups under rcu_read_lock(), while action objects may be freed under different locks without waiting for an RCU grace period.
That creates a race window where a freed tc_action object can be reclaimed before its reference count is checked, opening the door to controlled memory corruption.
The vulnerable path was ultimately reached through RTM_NEWTFILTER and RTM_DELTFILTER netlink operations rather than RTM_NEWACTION, which requires stronger privileges.
To make exploitation possible, the attack relies on unprivileged user namespaces being enabled, allowing an attacker to gain CAP_NET_ADMIN within a separate namespace.
The exploit also depends on specific traffic control components, including the clsact queueing discipline and flower filters, to avoid kernel paths that would otherwise serialize access and make the race harder to hit.
Jia Jie said AI helped speed up several stages of the work, including bug discovery, generation of a KASAN proof-of-concept, and optimization of the race condition.
The final exploit used multiple binder threads and a deleter thread operating on the same action index, reducing race time from more than 15 minutes to roughly five seconds in testing.
From there, the exploit reclaimed freed kmalloc-256 objects using user_key_payload allocations triggered via KEYCTL_UPDATE, eventually achieving instruction pointer control and a reliable root escalation chain on CentOS Stream 9 desktop systems.
The post-exploitation path overwrote core_pattern, a well-known technique that causes the kernel to execute an attacker-controlled binary as the core dump handler with root privileges.
In repeated tests on CentOS 9, the exploit succeeded consistently, though runtime varied depending on CPU conditions and tuning.
Lee Jia Jie noted that the method was reliable on the target environment but still relied on hardcoded ROP gadget offsets, making portability across kernel builds more difficult.
The disclosure is notable not only because of the bug’s impact, but because it reinforces a growing trend: AI can dramatically shorten the path from code review to exploit development, even in a target as complex as the Linux kernel.
While Jia Jie cautioned that AI still has reasoning gaps and blind spots, the research suggests defenders should expect faster discovery of race conditions, memory safety flaws, and exploit primitives in mature open-source codebases.
Organizations running Linux desktops or servers with unprivileged user namespaces enabled should review kernel versions immediately and apply patches for CVE-2026-53264.
Security teams should also monitor for suspicious netlink activity and unexpected changes to core_pattern, which may indicate active exploitation attempts.
Cut SOC investigation blind spots and contain threats earlier to reduce response costs and business disruption with ANY.RUN.
