RackN

A 16-Year-Old Bug Lives in the Firmware of Most of Your Servers

If your servers run UEFI firmware built on EDK II, and most do, they carry a 16-year-old buffer overflow in their network boot code. A machine is only exposed if it PXE boots, and since PXE is on by default in most server firmware, a reboot that falls through the boot order can get it there. Secure Boot doesn’t help, because the bug fires before Secure Boot looks at anything.

The useful question is which machines network boot, and how fast you can find out.

The bug

CVE-2026-97631 is in UefiPxeBcDxe, the PXE driver in EDK II’s NetworkPkg. EDK II is the open source base that AMI, Insyde and Phoenix build firmware from, and Microsoft’s Project Mu carries the same code. It scores 9.5 under CVSS 4.0, and the flawed copy has been there since about 2010. This is the same PXE code PixieFail hit in 2024, and the fuzzing that followed didn’t catch this one.

During a PXE boot, the driver copies incoming ICMP error messages into a fixed 504-byte field. The length checks in both the IPv4 and IPv6 paths are wrong, so an oversized message writes past the field. The easy result is a hung boot. With an exploit tuned to the specific model and firmware build, the result is code running in the DXE phase with full firmware privileges.

The code runs before Secure Boot or measured boot verify or measure anything in the OS, while the boot process can still reach the TPM. That’s a path to things like breaking TPM-sealed disk encryption.

Alex Matrosov at Anthropic and Dan Holmes at Google found it and planned coordinated disclosure for January. A public fix pull request let it out early, the TianoCore advisory landed October 22, and Ars Technica picked it up.

How worried to be

This isn’t an internet drive-by. The attacker has to be on your network during a boot window and has to tune the exploit per platform. Luke Jarymowycz on our team called it an internal actor problem, and that’s right.

Ransomware crews and nation-state actors already live inside networks, waiting for openings like this. And most provisioning networks we see are flat, shared and trusted on the theory that nobody gets in there.

The fix itself is easy once it exists. There are no keys to revoke, just a firmware update. Today, though, the fix is only in EDK II master (merged October 15). It isn’t in a stable EDK II release, and OEM firmware will follow on each vendor’s own timeline. That gap is your time to get organized.

Working through it

This is how RackN customers are approaching it with Digital Rebar.

  1. Inventory. Find out which machines have PXE in the boot order and what firmware they run. Most teams we talk to can’t do this without logging into BMCs one at a time. Digital Rebar reads firmware versions and boot order out of band, with no reboots, which turns a multi-week survey into an afternoon.
  2. Turn PXE off where you don’t need it. Remove PXE from the boot order, and disable the UEFI network stack where the firmware allows it, on every machine that doesn’t net boot. This is the main mitigation available today, and it’s a fleet-wide BIOS and boot order change. Old hardware that will never see a patch needs a retirement date.
  3. Patch as fixes arrive. Dell, HPE, Lenovo and Supermicro will ship on different days. Run each update through the same pipeline as it lands, then rescan the machine to confirm the new version actually took.
  4. Advanced: PXE only during provisioning. For machines that do need network boot, nothing authenticates that path, so every unplanned PXE boot is an open window. Out-of-band control can enable PXE for planned provisioning cycles and disable it otherwise. It adds operational overhead, but it closes the window on the machines step 2 can’t.

What this looks like at scale

One customer runs about 100,000 machines with four admins. They rehearse firmware and OS changes in a lab against shared infrastructure as code, then roll them across their global sites. A firmware CVE goes into the same change queue as everything else.

Another customer reset 40,000 machines across their data centers in six weeks, including firmware updates and onboarding new OEMs, on a project they’d started planning two months earlier.

Both got there by building the process once and running it continuously.

It won’t be the last one

This joins PixieFail, LogoFail, PKFail and BlackLotus. HD Moore expects it to be a “forever bug,” living on in end-of-support systems. With AI now helping researchers find flaws faster, firmware CVEs will keep coming.

So fix this one, and come out of it with a firmware cadence that makes the next one routine.

How long would it take your team to list which servers PXE boot and what firmware they run? If the answer is weeks, let’s talk. Digital Rebar is a low-risk way to get that process running, and you can build on it from there.