On 2026-09-28 Debian published DSA-6527-1 for rsync. It moves Debian 13 (trixie) to upstream rsync 3.5.0 and fixes 33 CVEs. According to the advisory, the bugs can lead to local privilege escalation, bypass of access restrictions, information disclosure, denial of service or arbitrary code execution. rsync is common on Debian servers, but how urgent the rsync DSA-6527-1 patch is depends on how rsync runs on that host. This post covers how to tell whether a box runs the rsync daemon, accepts rsync over SSH or only uses the client. It then walks through patching, checking that the new binary is the one actually running, testing backup jobs and tightening rsyncd.conf.
What happened: rsync 3.5.0 and DSA-6527-1
Upstream released rsync 3.5.0 on 13 Aug 2026. The release came out of audits of path handling and of the daemon protocol. Debian's advisory, signed by Salvatore Bonaccorso, lists these IDs:
- CVE-2026-53783 to CVE-2026-53786 (CVE-2026-53787 is not part of the advisory)
- CVE-2026-53788 to CVE-2026-53803
- CVE-2026-70452 to CVE-2026-70464
That makes 33 CVEs in total, which matches the count in the upstream NEWS file. The most severe is CVE-2026-53791. When a daemon has proxy protocol = true, a client that connects directly can send its own PROXY header with a forged source address and get past hosts allow/hosts deny. NVD published the CVE on 13 Aug 2026 and shows a secondary (CNA) CVSS 3.1 score of 9.1 Critical (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:N). I did not check the CVSS scores of the other 32 CVEs one by one, so I don't quote them here.
Based on the upstream NEWS, the fixes fall into these groups:
- Daemon access control: the PROXY spoofing above.
hosts denyfailed open when a hostname in the list could not be resolved (CVE-2026-70452).auth usersparsing was broken (CVE-2026-70463). The name converter mapped unknown names to uid/gid 0 (CVE-2026-53798). - Memory corruption a peer can trigger: out-of-bounds writes in
read_args()(CVE-2026-70456),add_implied_include()(CVE-2026-70461) and hard-link handling (CVE-2026-70458). - Daemon DoS: unauthenticated peers could stall connections indefinitely (CVE-2026-70464), clients could request unlimited Zstandard threads (CVE-2026-70455), and equal weak-checksum chains caused quadratic CPU use (CVE-2026-70453).
- Symlink races and path escapes on both sender and receiver. One example is the module-root escape when
use chroot = no(CVE-2026-53784). - The
rrsyncwrapper: a TOCTOU escape from the restricted directory (CVE-2026-53783). - Client side: command injection through unquoted substitutions in
RSYNC_CONNECT_PROG, daemon hooks and remote-shell arguments (CVE-2026-53790).rsync-ssldid not verify TLS (CVE-2026-70454).
Who is affected: Debian package status
The versions below come from the Debian security tracker (DSA-6527-1, DLA-4888-1 and the CVE-2026-53791 entry). Compare your installed version against the fixed version:
| Release | Vulnerable version | Fixed version | Status |
|---|---|---|---|
| Debian 12 bookworm | 3.2.7-1+deb12u6 | 3.5.0+ds1-0+deb12u1 | fixed via DLA-4888-1 |
| Debian 13 trixie | 3.4.1+ds1-5+deb13u4 | 3.5.0+ds1-0+deb13u1 | fixed via DSA-6527-1 |
| forky (testing) | - | 3.5.0+ds1-1 | fixed |
| sid | - | 3.5.0+ds1-1 | fixed |
Bookworm is now handled by Debian LTS, so its fix came as a Debian LTS advisory, DLA-4888-1, and not as a DSA. It also moves bookworm to rsync 3.5.0. The tracker lists the same 33 CVEs for DLA-4888-1 plus CVE-2026-119741. If you still run Debian 12, patch it the same way as trixie below; apt policy rsync shows whether your mirror already offers the fixed package.
That extra CVE matters for trixie. CVE-2026-119741 is a denial of service in the sender's hash_search(): an unauthenticated client pulling from a public daemon module can send colliding weak checksums and burn the sender's CPU across concurrent connections. Upstream fixed it in rsync 3.5.2. The tracker still lists trixie-security 3.5.0+ds1-0+deb13u1 as vulnerable and marks it no-dsa (minor issue), so there will be no separate DSA. The fix is in 3.5.0+ds1-0+deb13u2 in trixie-proposed-updates and should arrive with the next point release. Until then, a patched trixie daemon is less protected than a patched bookworm one.
Daemon, SSH server or client only: how urgent is the rsync patch?
rsync has three roles, and the exposure is different for each:
| Role | What it looks like | Who can reach it | Urgency |
|---|---|---|---|
| Daemon (rsyncd) | rsync --daemon, TCP 873, /etc/rsyncd.conf | anyone who can reach the port, often without authentication | patch today |
| Server over SSH | rsync --server started by sshd, often forced with command= or rrsync | holders of a valid SSH key | high, especially for restricted keys |
| Client only | cron or timer jobs running rsync src host:dst | the remote end you connect to | normal; patch at the next window |
Check the installed version
dpkg-query -W -f='${Package} ${Version}\n' rsync
apt-cache policy rsync
Anything below 3.5.0+ds1-0+deb13u1 on trixie or below 3.5.0+ds1-0+deb12u1 on bookworm is vulnerable to the 33 CVEs from the advisory. On trixie, 3.5.0+ds1-0+deb13u1 is still open to CVE-2026-119741. If you run a daemon with modules anyone can pull from, restrict port 873 to known clients (see the firewall rule below) or watch for 3.5.0+ds1-0+deb13u2 to show up in apt policy rsync.
Is the daemon running?
On Debian 13 the package ships /usr/lib/systemd/system/rsync.service. The unit has ConditionPathExists=/etc/rsyncd.conf, so it does nothing until that file exists. Check the unit, the config file and the port:
systemctl is-enabled rsync
systemctl status rsync --no-pager
ls -l /etc/rsyncd.conf
ss -tlnp 'sport = :873'
If ss shows port 873 owned by inetd, xinetd or systemd (socket activation) and not by rsync, the daemon is still in use. It is just started a different way. Also look for daemons started by hand or from scripts:
ps -eo pid,user,lstart,args | grep '[r]sync --daemon'
Is rsync accepting connections over SSH?
Forced commands in authorized_keys are the usual sign of a backup target:
grep -Hn -e 'rsync --server' -e 'rrsync' /root/.ssh/authorized_keys /home/*/.ssh/authorized_keys 2>/dev/null
grep -rn -e 'ForceCommand' -e 'AuthorizedKeysFile' /etc/ssh/sshd_config /etc/ssh/sshd_config.d/
Keys without a forced command can run rsync too. Check whether rsync server processes show up while your backup windows are running:
ps -eo pid,user,lstart,args | grep '[r]sync --server'
If none of these checks finds anything, the host probably only uses the rsync client. Patch it anyway. CVE-2026-53790 (command injection) and CVE-2026-70462 (client timeout overflow) both affect the client side, and a compromised backup target can exploit a client that connects to it.
Patch Debian 13 and verify the running binary
Warning: restarting the daemon drops any transfers in progress. Do it outside your backup window.
apt-get update
apt-get install --only-upgrade rsync
dpkg-query -W -f='${Version}\n' rsync
grep ' rsync:' /var/log/dpkg.log | tail -n 3
The daemon forks one child per connection from the long-running parent process. If the parent was started before the upgrade, it keeps running the old, deleted binary. Restart it and confirm:
systemctl restart rsync
pid=$(systemctl show -p MainPID --value rsync)
ls -l /proc/"$pid"/exe
A parent that is still running the old binary looks like this:
/proc/2711/exe -> /usr/bin/rsync (deleted)
After the restart the (deleted) suffix has to be gone. To catch every rsync process that is still on the old binary, including SSH-spawned servers and long client runs:
for p in $(pgrep -x rsync); do ls -l /proc/"$p"/exe; done
Each SSH-mode rsync --server is a new process, so new connections use the patched binary without any restart. Long-running transfers that started before the upgrade keep using the old code until they finish.
Check backup jobs after the upgrade
3.5.0 changes some behaviour on purpose, and that can break existing jobs. These are the items in the upstream NEWS most likely to affect backups:
- A non-daemon receiver now follows a symlinked destination directory only if root or the running user owns the symlink. Destinations symlinked by another uid are refused.
--insecure-linksbrings back the old behaviour, but fix the ownership instead. rrsyncin a restricted subdirectory now forces--no-Dand refuses--copy-unsafe-links.--compress-threadsis capped at 8 on a daemon.--max-alloc=0is rejected by both client and daemon in 3.5.0, and the trixie package (3.5.0+ds1-0+deb13u1) still rejects it, so remove it from your jobs. Upstream 3.5.1 accepts--max-alloc=0again, but only as another way to write the parser's maximum allocation limit. It no longer means "no limit".hosts denyentries that cannot be resolved now block the connection instead of letting it through. A stale hostname in that list can lock out more clients than before.proxy protocol = truewithoutproxy protocol hostsnow rejects every connection, and the daemon warns about it at startup.
Find every scheduled rsync job:
systemctl list-timers --all --no-pager
grep -rn rsync /etc/crontab /etc/cron.d/ /etc/cron.daily/ /var/spool/cron/crontabs/ 2>/dev/null
grep -rln rsync /etc/systemd/system/ /usr/local/bin/ 2>/dev/null
Do a dry run of each job by hand with the same options, and check the exit code:
rsync -a --dry-run --itemize-changes /srv/data/ backup@backup.example.net:/srv/backup/data/
echo "exit code: $?"
After the next scheduled run, check the logs for refused paths or failed units:
journalctl -u rsync --since today --no-pager
systemctl --failed --no-pager
If rsync copies are your only backup, this is a good time to add a second, versioned tool. The hardened restic setup with an append-only repository gives you snapshots that a compromised client cannot overwrite.
Harden rsyncd
If you don't need the daemon, turn it off. This stops anything that pulls from port 873:
systemctl disable --now rsync
If you do need it, run transfers as a dedicated unprivileged user, chroot every module, keep modules read-only and limit which hosts can connect. Warning: if you already run rsyncd, back up your existing /etc/rsyncd.conf and /etc/rsyncd.secrets before you change anything. Replacing the secrets file with an empty one locks out every client that uses auth users after the next restart. The commands below only create the secrets file if it does not exist yet:
[ -e /etc/rsyncd.conf ] && cp -a /etc/rsyncd.conf /etc/rsyncd.conf.bak
[ -e /etc/rsyncd.secrets ] && cp -a /etc/rsyncd.secrets /etc/rsyncd.secrets.bak
useradd --system --no-create-home --shell /usr/sbin/nologin rsyncd
[ -e /etc/rsyncd.secrets ] || install -m 600 -o root -g root /dev/null /etc/rsyncd.secrets
uid = rsyncd
gid = rsyncd
use chroot = yes
read only = yes
list = no
reverse lookup = no
max connections = 4
timeout = 600
[mirror]
path = /srv/mirror
comment = read-only mirror for the backup host
hosts allow = 192.0.2.10 198.51.100.0/24
auth users = backupuser
secrets file = /etc/rsyncd.secrets
Notes on this config:
use chroot = yesmakes chroot mandatory. With the default, the daemon tries to chroot but keeps going if that fails. CVE-2026-53784 was exactly a module escape when chroot was not in use.hosts allowmatches client IPs. Withreverse lookup = no, use IPs and CIDRs, not hostnames. That also avoids the unresolvable-name problem fixed by CVE-2026-70452.timeoutandmax connectionsare both unlimited (0) by default. Setting them limits the damage from stalled or flooding clients.- Don't enable
proxy protocolunless a load balancer sits in front of the daemon. If you do enable it, also setproxy protocol hosts(see the 3.5.0 section of the upstream NEWS; the trixiersyncd.conf(5)manpage does not document this parameter). - The secrets file holds
backupuser:passwordlines.strict modesis on by default and rejects a world-readable file.
On a new setup the secrets file is still empty and the module path may not exist yet. Create the directory and add the user line (pick your own password; the file keeps mode 600):
mkdir -p /srv/mirror
printf 'backupuser:%s\n' 'CHANGE-THIS-PASSWORD' >> /etc/rsyncd.secrets
chmod 600 /etc/rsyncd.secrets
ls -l /etc/rsyncd.secrets
Restart the daemon (this drops active transfers) and test from an allowed client:
systemctl restart rsync
rsync --list-only rsync://backupuser@backup.example.net/mirror/
The client asks for the password interactively. For scripts, set RSYNC_PASSWORD or put the password alone in a file readable only by the job's user and pass it with --password-file.
On top of that, filter TCP 873 so only the backup hosts can reach it. A fragment for the default inet filter input chain:
tcp dport 873 ip saddr { 192.0.2.10, 198.51.100.0/24 } accept
tcp dport 873 drop
Debian's unit already sets ProtectSystem=full, PrivateDevices=on and NoNewPrivileges=on. You can restrict it further with a drop-in, as described in sandboxing systemd services with systemd-analyze.
For SSH targets, give every backup key a forced command restricted to one directory, for example command="rrsync -ro /srv/backup". Combine that with the key restrictions from hardening SSH on Debian 13. Keep in mind that rrsync was itself affected (CVE-2026-53783). A forced command is a second layer and does not replace the patch.
What to watch next
- Bookworm: Debian LTS fixed these CVEs with DLA-4888-1 (
3.5.0+ds1-0+deb12u1). The same 3.5.0 behaviour changes apply there, so test bookworm backup jobs just like trixie ones. - CVE-2026-119741 on trixie: the
hash_search()CPU-exhaustion DoS is still open in3.5.0+ds1-0+deb13u1(no-dsa). The fix,3.5.0+ds1-0+deb13u2, sits in trixie-proposed-updates and is expected with the next point release. Until it lands, keep port 873 restricted and checkapt policy rsyncafter each point release. - Regressions: upstream released 3.5.1 on 21 Sep 2026 with fixes for 3.5.0 regressions, and forky/sid already have 3.5.2 (
3.5.2+ds1-2). The trixie package3.5.0+ds1-0+deb13u1already includes six patches backported from 3.5.1, according to its Debian changelog. They fix "Permission denied" on paths through directories with mode 0711 or 0111,rrsyncrefusing/and resolving paths like--link-dest=/previousrelative to the destination (which silently turned incremental backups into full copies),--files-fromlists outside--confine-rootbeing refused for local and remote-shell transfers,--relativeor--files-fromsources through a symlink failing with "Too many levels of symbolic links", and a symlink being followed as the--link-dest,--copy-destor--compare-destbasis file. If your trixie jobs fail in some other way after the upgrade, look it up in the Debian bug reports for rsync and the security tracker before you change your jobs. - Logs: after the upgrade, more connections refused by
hosts denyor more refused symlinked destinations in the logs come from the stricter checks. They can also show you a misconfiguration you didn't know about.
Takeaways
- On trixie, upgrade to
rsync 3.5.0+ds1-0+deb13u1. DSA-6527-1 fixes 33 CVEs. - Use
ss -tlnp 'sport = :873',systemctl status rsyncand a grep ofauthorized_keysto find out whether a host runs the daemon, accepts SSH transfers or only uses the client. - Restart
rsync.serviceand check that/proc/PID/exeno longer shows(deleted). - Dry-run every cron job and timer that uses rsync. Watch for refused symlinked destinations and stricter
rrsyncrules. - Configure rsyncd with
use chroot = yes, a dedicateduid/gid,read only = yes,hosts allowand a firewall rule on port 873. If you don't need the daemon, disable it. - On bookworm, upgrade to
rsync 3.5.0+ds1-0+deb12u1from DLA-4888-1.
Sources
- DSA-6527-1 rsync security update (debian-security-announce, 2026-09-28)
- Debian security tracker: DSA-6527-1
- Debian security tracker: DLA-4888-1
- Debian security tracker: CVE-2026-53791
- Debian security tracker: CVE-2026-119741
- Debian security tracker: rsync source package
- NVD: CVE-2026-53791
- rsync NEWS (3.5.0 and 3.5.1 release notes, incl. proxy protocol hosts and --max-alloc=0)
- Debian sources: rsync debian/changelog (3.5.0+ds1-0+deb13u2, incl. deb13u1 backports from 3.5.1)
- Debian packages: rsync in trixie
- Debian packages: rsync trixie amd64 file list
- Debian sources: rsync.service (3.5.0+ds1-0+deb13u2)
- rsyncd.conf(5) - Debian trixie manpage
- rrsync(1) - Debian trixie manpage
Comments