A normal `apt upgrade` on Debian 12 or Ubuntu 22.04 or later can install a package nobody signed, as long as someone controls the network between the machine and the mirror. The flaw is CVE-2026-109518, scores 7.7 out of 10 on CVSS v4 and the apt maintainers consider it critical because the result is code executed as root. Canonical and Debian published the fixed packages on November 4.
The Ubuntu advisory is signed by the official CVE-2026-109518 entry for APT and the Debian fix ships in DSA-6607-1, with specific versions for bookworm and trixie. What makes the case interesting is the condition that triggers everything: client retries. APT has been configured with `Acquire::Retries` set to 3 since Debian 12 and Ubuntu 22.04 LTS, a setting meant to keep a download from failing when the mirror hiccups. It is precisely that retry that opens the door.
Anyone in the middle can let the package be verified and change the content right afterwards, when the client requests it again. The package's maintenance scripts run as root without anyone rechecking the signature. The result is an installation that looks normal in the logs and leaves the system in the hands of whoever controlled the wire.
The mechanism, plainly put
APT downloads a package, checks its hash against the signed index and leaves it ready to install. If the download is cut off, the retry requests the file again, and if the server responds with a redirect whose destination already exists in the local cache, APT reuses the previous file to avoid downloading it again.
There lies the bug: two different downloads share the same hard link node and the second can write over the already verified content. Max Justicz discovered the flaw and the Debian tracking description sums it up in two vectors: injection via redirect and replacement through reused hard links.
The QSB-122 advisory from Qubes OS exists because its Debian and Whonix templates inherit the problem. There the practical detail appears: on Qubes the repositories use HTTPS by default, so the second vector disappears and the first remains alive if someone changed the configuration by hand.
The useful conclusion is that HTTPS reduces the risk, but does not fix the flaw. Only the new package does, and the proof is in the installed version, not in the final message of the upgrade.
Who is really exposed?
Debian 12 and 13, Ubuntu 22.04, 24.04, 26.04 and 26.10, and any derivative that uses apt with retries enabled. Older versions are outside the default risk because they ship `Acquire::Retries` set to zero, although any Dockerfile or CI script that has raised it puts them back in the affected group.
That is the point most often overlooked in internal discussions. Continuous integration teams and container images are exactly where someone bumps the retry count so the build doesn't fail against a slow mirror.
The AWS advisory treats it as a fleet problem rather than a desktop one: instances, container images and deep learning AMIs, with updated AMIs scheduled for November 13 and ParallelCluster 3.16.3 for the 16th.
The numbers to look for are specific. Debian 12 needs apt 2.6.2 and Debian 13, 3.0.4. Ubuntu 22.04 is fixed with 2.4.15ubuntu0.1, 24.04 with 2.8.4ubuntu0.1, 26.04 with 3.2.1ubuntu0.1 and 26.10 with 3.3.3ubuntu0.1. Versions 20.04 and earlier are patched through Ubuntu Pro. The same kind of verification error has already appeared in the CVE-2026-98432 in the Linux kernel, where Ubuntu and RHEL were also slow to publish a package.
Update your system
Here is the catch that is worth reading twice: the command that installs the patched apt runs with the vulnerable apt. Canonical recommends disabling retries for that specific update, and that order it matters more than it seems.
In practice it comes down to two commands: `apt update` and then `apt upgrade -o Acquire::Retries=0`, repeated on every host that does not come from an already patched image. Then check the version with `apt –version`, because a successful upgrade message does not say which version ended up installed. That order is not a matter of style: the update that closes the flaw runs while apt is still vulnerable.
The second step is basic hygiene and applies forever: move repositories to HTTPS wherever HTTP remains. The same lesson left by the OpenSSH flaw on musl systems applies here: the updated package is not enough if the service keeps running the old binary. It does not replace the patch, but it removes the easiest vector and prevents the next flaw of this kind from finding the same door open.
The third is boring and necessary: review the logs from the last few days looking for installations nobody scheduled. A package installed outside the maintenance window, or two installations of the same package minutes apart, are the typical fingerprint of a replacement during a retry.
What this flaw teaches about the supply chain
The supply chain is usually told as a story of compromised dependencies and stolen maintainer accounts. This case is more uncomfortable because there is no malicious third party in the chain: there is a reasonable retry function that, combined with a redirect, breaks the guarantee that what was verified is what gets installed.
Systems that survive that kind of failure are the ones that verify twice and treat the cache as untrusted material. APT did this reasonably well until the hard-link shortcut saved a download, and that shortcut is exactly what the patch removes in Debian versions 2.6.2 and 3.0.4 and in the four Ubuntu builds.
The lesson for anyone managing fleets is not to distrust apt, but to remember that verification and installation are two distinct moments. Anything that happens between them is attack surface, even if the process looks internal.
Pending tasks
Neither Debian nor Ubuntu reported exploitation at the time the advisories were published, and that is good news with an expiration date: the flaw is silent, leaves no visible error and fits into any internal network with HTTP mirrors. Checking that no one used it means looking at what was installed, where it was downloaded from and whether the configured mirror is still the one the systems team believes it is.
For anyone managing fleets, this week's work is inventory: which hosts are still on the old version, which container images need rebuilding and which CI scripts have the retry bumped up. The right answer is not to turn off retries forever, but to set them to zero while apt is updated.