Critical DHCPv6 Vulnerability in OpenWrt Opens the Door to Code Execution with Root Privileges

The OpenWrt project has released update 24.10.8, fixing a critical vulnerability in its DHCPv6 implementation. The flaw, assigned the identifier CVE-2026-53921 and rated 9.8 on the CVSS 3.1 scale in the project's GitHub advisory, allows an unauthenticated attacker with network access to a DHCPv6 server to overwrite a buffer on the stack in the odhcpd daemon using a specially crafted DHCPv6 REQUEST.

The danger is compounded by the fact that odhcpd runs with root privileges, while embedded hardware, as the advisory notes, typically lacks stack canaries and address space layout randomization (ASLR). On typical devices this makes arbitrary code execution a thoroughly realistic scenario — meaning a complete takeover of the router, not merely a crash.

A packet to port 547

The advisory describes two independent overflow sites along the request-handling path. In both cases, specially crafted IA options leave insufficient room in the fixed 512-byte buffer, after which the code appends the response data without proper bounds checking. The final trigger is an unauthenticated REQUEST sent to UDP port 547. The advisory comes with public Python code demonstrating exploitation of both paths.

Users on the 24.10 branch are advised to install 24.10.8, and users on 25.12 to install version 25.12.5. As of July 28, no signs of exploitation in real-world attacks have been recorded, and the vulnerability is absent from the CISA KEV catalog, though this is not yet a guarantee that no attacks are taking place. The 24.10 branch is in its security-support phase, scheduled to end in September 2026; the project recommends moving to the 25.12 series.

The same release also closes other pre-authentication weaknesses in odhcpd — an out-of-bounds write, a use-after-free, a memory leak and others — as well as three request smuggling bugs in uhttpd.

The role of AI in discovery and fixes

A separate audit by the company Hacker House, using AI, uncovered command injection, directory traversal and XSS vulnerabilities in optional LuCI components; those fixes are still under review. The company described its method as a four-stage “inference-fuzzing”: the broad-sweep stage uses the Qwen model, while precise filtering of the results uses a frontier model such as Anthropic Claude Opus 4.6, after which the findings are verified manually by researchers.

Interestingly, OpenWrt itself also used AI in preparing the fixes: some commits are marked with an “Assisted-by: Claude” trailer, and an automated pull request review was generated with Claude Code.

Source: The Hacker News