high

CVE

CVE-2026-124757

CWE

CWE-78, CWE-863, CWE-150

Affected Surface

  • Debian 12 bookworm, needrestart 3.6-4+deb12u3 (local root confirmed by the reporter)
  • Ubuntu 24.04 LTS, needrestart 3.6-7ubuntu4.5 (local root confirmed by the reporter). Ubuntu lists 16.04, 18.04, 20.04, 22.04, 24.04, 26.04 and 26.10 as vulnerable
  • Debian 13 trixie, forky and sid, needrestart 3.11-1 (session spoof and root notify-send with attacker-chosen DISPLAY and text; command execution not reproduced)
  • Upstream needrestart before 3.11.1
  • Exploitation of the notify hook needs /usr/bin/notify-send (libnotify-bin) on the host and a root run of needrestart, which the stock apt hook triggers after every apt or dpkg transaction

TL;DR

needrestart checks which running processes still use outdated libraries after a package upgrade. On Debian and Ubuntu its apt hook runs it as root after every apt or dpkg transaction, and it can notify logged-in users through hooks in /etc/needrestart/notify.d/.

Armaan Sandhu found that two of its assumptions combine into local root (CVE-2026-124757, assigned by Red Hat, disclosed 17 November 2026):

  1. needrestart decides which user and session a process belongs to by matching unanchored regexes against /proc/PID/cgroup. On a systemd host a user can create subdirectories inside their own delegated user@UID.service cgroup. A directory named user-0.slice/session-1.scope makes their process look like part of root’s login session.
  2. The packaged 400-notify-send hook puts the list of process names into a su -c shell string. A process picks its own name with prctl(PR_SET_NAME) (15 bytes). A name that closes the quote and adds shell syntax is executed by the shell su starts, and with the forged session that shell runs as root.

The reporter confirmed root on Debian 12 bookworm (3.6-4+deb12u3) and Ubuntu 24.04 (3.6-7ubuntu4.5). On Debian 13 trixie (3.11-1) the session spoof works and notify-send runs as root with attacker-chosen DISPLAY and text, but command execution wasn’t reproduced. The attacker waits for the next root apt run, which includes routine upgrades by an admin. The reporter didn’t test unattended-upgrades.

The hook exits early unless /usr/bin/notify-send (package libnotify-bin) is installed, so hosts without it are not exposed to the root-shell path. They still get the session misclassification and the terminal-escape issue fixed alongside it.

What to do now:

  1. Install your distribution’s fixed needrestart as soon as it ships. On 18 November neither Debian nor Ubuntu had one. Upstream fixed it in needrestart 3.11.1.
  2. Until then, turn off the notify-send hook: add NR_NOTIFYD_DISABLE_NOTIFY_SEND='1' to /etc/needrestart/notify.conf, or set $nrconf{sendnotify} = 0; in a file under /etc/needrestart/conf.d/.
  3. On multi-user hosts and shared build machines, treat this as urgent: any local account, including a compromised service account, can wait for the next root apt run.

Affected packages and projects

EcosystemPackageAffected versionsFixed or safe version
Upstreamneedrestartbefore 3.11.13.11.1 (tagged 17 Nov 2026 20:30 UTC) and master after e8ae1f2
Debian 12 bookwormneedrestart3.6-4+deb12u3 (root confirmed)not fixed as of 18 Nov. Upstream says the second patch needs a backport to 3.6
Debian 13 trixie, forky, sidneedrestart3.11-1 (session spoof, root notify-send)not fixed as of 18 Nov
Ubuntu 24.04 LTS nobleneedrestart3.6-7ubuntu4.5 (root confirmed)not fixed as of 18 Nov
Ubuntu 16.04, 18.04, 20.04, 22.04, 26.04, 26.10needrestartlisted as vulnerable by Ubuntunot fixed as of 18 Nov

The reporter didn’t check other distributions. Upstream notes the notify hook runs whenever needrestart runs, so hosts that call needrestart from monitoring or configuration management, not only from apt, are exposed too.

Technical analysis

1. Filing a process under root’s session

needrestart reads each process’s cgroup path and runs it through a chain of regexes to work out the owning uid and session. From needrestart before the fix:

# `systemd --user` manager
if($cgroup =~ m@/user\@(\d+)\.service/init\.scope$@ && $ptable->{$pid}->{fname} ne "(sd-pam)") {
    ...
}

if($cgroup =~ m@/user-(\d+)\.slice/session-(\d+)\.scope@) {          # <-- no ^ anchor
    print STDERR "$LOGPREF #$pid part of user session: uid=$1 sess=$2\n" if($nrconf{verbosity} > 1);
    push(@{ $sessions{$1}->{"session #$2"}->{ $ptable->{$pid}->{fname} } }, $pid);
    next;
}
if($cgroup =~ m@/user\@(\d+)\.service(/.+\.slice)*/([^/]+\.service)$@) { ... }
if($cgroup =~ m@/user\@(\d+)\.service@) { ... }
  • systemd delegates each user’s user@UID.service subtree to that user, so they can create cgroup directories under it and move their own processes into them.
  • A real user service lives at something like /user.slice/user-1000.slice/user@1000.service/app.slice/foo.service. That path doesn’t contain /user-N.slice/session-N.scope back to back, so the session regex doesn’t match it.
  • If the user creates .../user@1000.service/app.slice/x/user-0.slice/session-1.scope and moves a process there, the unanchored session regex finds /user-0.slice/session-1.scope in the middle of the path and records the process under uid 0, session #1. It never reaches the user@1000.service checks below.

The fix anchors all four matches so they only accept systemd’s slice components from the start of the path:

-		if($cgroup =~ m@/user\@(\d+)\.service/init\.scope$@ && $ptable->{$pid}->{fname} ne "(sd-pam)") {
+		if($cgroup =~ m@^(?:/[^/]+\.slice)*/user\@(\d+)\.service/init\.scope$@ && $ptable->{$pid}->{fname} ne "(sd-pam)") {
 ...
-		if($cgroup =~ m@/user-(\d+)\.slice/session-(\d+)\.scope@) {
+		if($cgroup =~ m@^(?:/[^/]+\.slice)*/user-(\d+)\.slice/session-(\d+)\.scope@) {
 ...
-		if($cgroup =~ m@/user\@(\d+)\.service(/.+\.slice)*/([^/]+\.service)$@) {
+		if($cgroup =~ m@^(?:/[^/]+\.slice)*/user\@(\d+)\.service(/.+\.slice)*/([^/]+\.service)$@) {
 ...
-		if($cgroup =~ m@/user\@(\d+)\.service@) {
+		if($cgroup =~ m@^(?:/[^/]+\.slice)*/user\@(\d+)\.service@) {

^(?:/[^/]+\.slice)* allows only .slice path components before the user slice or service, so text buried inside a user-owned subtree no longer matches (7874c08). The commit message notes this was tested against systemd layouts only, and container setups with a different cgroup prefix may now be classified differently.

2. Turning a process name into a root command

For each session with stale processes, needrestart runs the notify hooks with NR_USERNAME, NR_SESSION and NR_SESSPPID set, and passes the process list (names and PIDs) on stdin. The Debian and Ubuntu 3.6 hook (ex/notify.d/400-notify-send in upstream v3.6):

NSEND='/usr/bin/notify-send'
test -x "$NSEND" || exit 1                              # no libnotify-bin: hook does nothing

case "$NR_SESSION" in
  session*)
    DBUS_SESSION_BUS_ADDRESS=$(sed -z -n s/^DBUS_SESSION_BUS_ADDRESS=//p "/proc/$NR_SESSPPID/environ")
    export DISPLAY=$(sed -z -n s/^DISPLAY=//p "/proc/$NR_SESSPPID/environ")
    ...
    if [ -z "$DISPLAY" ]; then exit 1; fi              # attacker sets DISPLAY in their own environ

    MSGTITLE=$(gettext 'Relogin or restarts required!')
    MSGBODY=$(gettext '...')'\n'$(cat)                 # stdin: "<process name>[<pid>]" list

    su -p -s /bin/sh -c "$NSEND -a needrestart -u critical -i dialog-warning \"$MSGTITLE\" \"$MSGBODY\"" "$NR_USERNAME"
    ;;
esac
  • NR_USERNAME is root because of the forged session, so su ... root starts /bin/sh as root.
  • $MSGBODY is expanded by the outer shell into the -c string before su runs. The inner root shell then parses it again. A process name containing " followed by shell syntax breaks out of the quoted argument.
  • The DISPLAY check reads the environment of the session’s process, which the attacker controls, so it doesn’t stop anything.

The fix passes the text through the environment so the inner shell only ever expands it as data (b1996a1):

-    MSGTITLE=$(gettext 'Relogin or restarts required!')
-    MSGBODY=$(gettext 'Your session is running obsolete binaries or libraries as listed below.
+    NR_NOTIFY_TITLE=$(gettext 'Relogin or restarts required!')
+    NR_NOTIFY_BODY=$(gettext 'Your session is running obsolete binaries or libraries as listed below.
 <i><b>Please consider a relogin or restart of the affected processes!</b></i>')'\n'$(cat)
 
-    su -p -s /bin/sh -c "$NSEND -a needrestart -u critical -i dialog-warning \"$MSGTITLE\" \"$MSGBODY\"" "$NR_USERNAME"
+    # Pass the text to su through the environment: the inner shell expands it as
+    # data, so process names in the body are never parsed as shell code.
+    export NR_NOTIFY_TITLE NR_NOTIFY_BODY
+
+    su -p -s /bin/sh -c "$NSEND -a needrestart -u critical -i dialog-warning \"\$NR_NOTIFY_TITLE\" \"\$NR_NOTIFY_BODY\"" "$NR_USERNAME"

The \$ escapes keep the variable names literal in the -c string. su -p preserves the environment, so the root shell substitutes the values inside double quotes, where they stay one argument each.

3. Escape sequences in process names

The third patch (e8ae1f2) replaces every byte outside printable ASCII with ? in displayed process names and in the text sent to hooks:

-	    push(@{$fnames{ $ptable->{$pid}->{fname} }}, $pid);
+	    push(@{$fnames{ $ptable->{$pid}->{fname} =~ s/[^\x20-\x7e]/?/gr }}, $pid);

Before it, a process name could carry terminal escape sequences into the admin’s apt output, or a newline that split a line of the machine-readable NEEDRESTART-PID output.

4. Why trixie behaves differently

On trixie’s 3.11-1 the reporter could still forge root’s session and get notify-send run as root with the attacker’s DISPLAY and text, but didn’t reproduce command execution. The upstream patch series applies to 3.11 with offsets, and the fix to the hook’s quoting is needed there too. Treat trixie as affected and patch it: a root notify-send talking to an attacker-chosen X display isn’t something to leave in place.

IOCs

This is a local privilege escalation with no malware or network component. Signs of an attempt:

TypeIndicatorNotes
cgroupDirectories named user-<N>.slice or session-<N>.scope nested inside any user@<UID>.service subtree under /sys/fs/cgrouplegitimate systemd layouts never put these below a user service
ProcessProcesses whose comm contains ", ;, $(, backticks, `or&`
Processsu -p -s /bin/sh -c /usr/bin/notify-send ... running as root with a body containing shell metacharactersvisible in process accounting or auditd execve records during apt runs
Logneedrestart verbose output reporting part of user session: uid=0 for a process owned by a non-root userrun needrestart -v -b to see classifications

Detection and remediation

# 1. Installed version
dpkg-query -W needrestart

# 2. Is the notify-send hook able to run? (no libnotify-bin = no root-shell path)
dpkg -s libnotify-bin 2>/dev/null | grep -E '^(Status|Version)'
ls -l /etc/needrestart/notify.d/400-notify-send

# 3. Look for forged session directories inside user service subtrees
find /sys/fs/cgroup -path '*/user@*.service/*' \( -name 'user-*.slice' -o -name 'session-*.scope' \) 2>/dev/null

# 4. Look for processes with shell metacharacters in their names
for p in /proc/[0-9]*; do c=$(cat "$p/comm" 2>/dev/null); case "$c" in *[\"\;\$\`\|\&]*) echo "${p#/proc/} $c";; esac; done

# 5. Past root shells spawned by the hook (if auditd records execve)
ausearch -m EXECVE -i 2>/dev/null | grep -E 'notify-send.*dialog-warning' | head

Mitigate now, then upgrade when your distribution ships a fix:

# Option A: disable only the notify-send hook
echo "NR_NOTIFYD_DISABLE_NOTIFY_SEND='1'" | sudo tee -a /etc/needrestart/notify.conf

# Option B: stop needrestart sending any user notifications
echo '$nrconf{sendnotify} = 0;' | sudo tee /etc/needrestart/conf.d/99-no-notify.conf

# Option C (desktop-free servers): remove libnotify-bin if nothing needs it
sudo apt-get remove libnotify-bin

# Later: take the fixed package
sudo apt-get update && sudo apt-get install --only-upgrade needrestart

Apply the mitigation before your next apt run on a multi-user host, since that run is the trigger. If you find forged cgroup directories or suspicious process names, treat the host as rooted: review /etc/sudoers*, /root/.ssh/authorized_keys, setuid binaries and systemd units changed since the last apt run.

Timeline

Date (UTC)Event
2026-10-23Armaan Sandhu reports the issue to needrestart upstream
2026-11-05Reported to the Debian and Ubuntu security teams
2026-11-08Reported to the linux-distros list
2026-11-17 12:15Upstream commits b1996a1, 7874c08 and e8ae1f2
2026-11-17 20:30Coordinated disclosure (21:30 CET). Upstream tags security release v3.11.1
2026-11-17 20:45Write-up posted to oss-security
2026-11-17Ubuntu publishes the CVE with priority High
2026-11-18Debian (bookworm, trixie, forky, sid) and Ubuntu (16.04 to 26.10) still list the package as vulnerable

From research to remediation

Check whether this pattern exists in your codebase

Turn this research into a remediation workflow. Scan dependencies and package manifests for similar supply-chain risk, then prioritize fixes with reachability context.

References