An 18‑Year‑Old Linux Kernel 0‑Day Fixed with Just Six Lines of Code—Discovered by AI

Corvus AI’s Birth

In 2026 AI could already write and analyze code, but turning a memory error into a stable exploit remained a complex, multi‑step problem. TencentOS’s security team built Corvus AI, a multi‑agent Harness framework that integrates model capabilities with professional security tools and experiment‑state management to keep long‑chain kernel research coherent and verifiable.

Why the Vulnerability Could Hide 18 Years

The flaw, named SCTPhantom , resides in the SCTP protocol stack. SCTP is used in telecom and specialized infrastructure and allows a connection to hold multiple network addresses that can be added or removed dynamically. The bug lies in the address‑deletion logic: when three specific parameters appear in a single SCTP message in a particular order, a kernel object that has already been freed is later accessed, creating a use‑after‑free condition.

This condition passes individual function checks and therefore evades typical static analysis and fuzzing. The vulnerable code entered the upstream tree in 2008 with v2.6.25 and survived hundreds of kernel releases, global code reviews, and extensive testing by tools such as Google syzkaller without being discovered.

From a Crash to Stable Root

Corvus AI autonomously generated a PoC that initially caused a kernel crash, proving the issue could be triggered. The agent then participated in root‑cause analysis, evaluating whether an attacker could control critical kernel objects, obtain necessary address information, and build a reliable exploit chain under real‑world defenses. After multiple experimental iterations and cross‑validation with security researchers, the AI produced a stable local‑privilege‑escalation exploit that gains root from an unprivileged user and, in certain container configurations, escapes to the host.

3 Hours to Migrate the Exploit

To assess the vulnerability’s impact on other distributions, the team gave the TencentOS exploit to Corvus AI for migration to Debian’s default kernel. The AI automatically analyzed differences between the two kernel environments, relocated and adjusted 29 kernel offsets and symbols, and continuously refined the code based on compilation and runtime results. The entire migration and verification process took roughly three hours.

Verification succeeded on:

TencentOS Server – local privilege escalation in default configuration.

Debian 13 – local privilege escalation in default configuration.

Ubuntu 24.04 – local privilege escalation in default configuration.

RHEL 9.8 / Rocky 9.8 – privilege escalation after loading the SCTP module.

Docker environments – the exploit also enables container escape to obtain host‑root.

These results demonstrate that SCTPhantom has been present in upstream code since v2.6.25 and affects a wide range of distributions.

This Wasn’t a Coincidence

The Linux kernel contains tens of millions of lines of code across networking, filesystems, virtualization, drivers, and permission management. Relying solely on manual target selection makes sustained vulnerability discovery infeasible. Corvus AI follows a full research pipeline: it continuously proposes hypotheses, runs experiments, analyzes outcomes, and iterates, building a persistent evidence chain until a stable exploit is achieved.

During the same period, Corvus AI independently discovered another local‑privilege‑escalation bug in the Open vSwitch module (CVE‑2026‑64531), confirming that the framework can uncover distinct vulnerabilities across different kernel subsystems.

From CVE Announcement to Pre‑Patch

Traditional OS‑vulnerability response starts after a public CVE appears. TencentOS’s team moves the starting point forward: once Corvus AI completes discovery and impact verification, the security team immediately begins impact assessment, patch development, kernel rebuild, and regression testing, allowing risk mitigation before a CVE number is assigned.

Timeline:

July 12 – Corvus AI discovers the bug and creates a KASAN PoC.

July 14 – SCTP subsystem maintainer acknowledges the issue.

July 15‑23 – Upstream kernel, TencentOS, Debian, Rocky/RHEL, and Ubuntu privilege‑escalation verification completed.

July 23 – Patch merged into Linux mainline (commit 9b2854f on GitHub).

July 27 – Initial validation of container‑escape scenario.

August 3 – Patch back‑ported to the Linux stable branch.

August 4 – CVE‑2026‑64564 officially assigned and TencentOS Server releases security advisory and updated packages.

The team follows responsible disclosure, coordinating with the Linux community and reporting to regulatory authorities, and refrains from publishing full exploit code until affected products are patched.

About Corvus AI

Corvus AI (Corvus AI) is an intelligent agent built by the TencentOS security team for Linux kernel vulnerability research. Leveraging a multi‑agent collaborative architecture and the Harness framework, it automates the entire workflow from bug discovery, root‑cause analysis, stable exploit construction, cross‑environment validation, to patch development, aiming to raise the overall security baseline of open‑source Linux distributions.