Skip to content

Zimbra Collaboration Suite CVE-2026-73570

Tags

Summary

Microsoft Threat Intelligence tracked exploitation of CVE-2026-73570, an unauthenticated OS command injection in Zimbra Collaboration Suite's SNMP notification path (full post text captured by this wiki via the RSS content:encoded field, Sep 30). A specially crafted SMTP request introduces untrusted input into SNMP notification processing; when a service-state change triggers health monitoring, swatchdog incorporates the attacker-controlled value into a snmptrap shell invocation → code execution as the zimbra service account. Preconditions: the optional zimbra-snmp package installed + SNMP notifications enabled. Fixed in Zimbra 10.1.20 (July 20, 2026); publicly disclosed August 13, 2026.

The exploitation window that matters: Microsoft observed active targeting of the same injection path between the July 20 fix and the August 13 disclosure, affected organizations in more than one region and industry; the composite spans multiple confirmed compromises (no single host necessarily shows every stage). The activity mixes automated payload delivery with hands-on-keyboard.

Pre-disclosure probing (Jul 28 - Aug 7)

Two out-of-band scanning tools probed the injection point using the same swatchdog→snmptrap path later used in exploitation, validating execution with OOB callbacks to unique subdomains on oast[.]fun, oast[.]online, dnslog[.]pp[.]ua, requestrepo[.]com, and campaign infrastructure under bypass[.]eu[.]org. Tells: the CVE-specific ZB73570 User-Agent for HTTP probes; randomized callback subdomains for DNS/ICMP; curl/wget/ping/nslookup/id identity checks; a small system-fingerprint script dropped in the webroot to prove external reach. Durable: OOB validator fleets that name the CVE in a User-Agent before disclosure = exploitation staging; the collaborator-subdomain hunt rule already on-wiki (oastify/interactsh) extends to oast.online and dnslog.pp.ua.

Post-exploitation tradecraft (the durable part)

Web shells: change webroot permissions → reconstruct an encoded/compressed payload from staged fragments → write JSP webshells into Jetty/mailboxd application paths, with copies on peer mailbox nodes; fragments then deleted; some cases temporarily opened a public directory, deployed, and restored the original permissions (permission-diff hunting defeated).

Privilege escalation, the marquee technique: abuse of Zimbra's own sudo-authorized service helpers: 1. Verify the helpers exist and the zmmailboxdmgr log directory is writable. 2. Back up /etc/pam.d/sudo; replace the legitimate zmmailboxd.out log with a symlink to the PAM config; invoke the privileged zmmailboxdmgr → the PAM file becomes zimbra-owned. 3. Add a pam_exec session hook; trigger a sudo session through the legitimate zmstat-fd helper → hook runs as root → NOPASSWD: ALL sudoers entry for zimbra. 4. Restore original PAM content, remove staging artifacts, keep the sudoers entry. Confirmed by subsequent passwordless sudo. Fallback rails in the controller: Postfix, a Java agent, zmstat-fd, nginx. Read: a privileged helper writing to an attacker-writable path, following symlinks, is a root-delivery primitive; the restore-the-original-file step means config-integrity checks pass while the sudoers drop persists.

Host persistence: zimlog.service, a systemd unit manually installed to /etc/systemd/system/, named like a Zimbra logging component, moved with sudo, owned by the service account, timestamps modified to match rsync.service/sshd.service, enabled at boot.

Cluster lateral movement: the actor mapped the deployment with zmprov (mailbox/MTA roles), tested for /opt/zimbra/.ssh/zimbra_identity (the cluster's built-in SSH trust), then SSH'd noninteractive with host-key verification disabled and moved payload fragments/webshells with rsync, validating receipt and deleting sources. The trusted cluster identity is the movement rail, one edge node compromises the mail cluster.

Credential and mailbox collection: zmlocalconfig -s exposes LDAP/MySQL/Postfix/Amavis/replication credentials; authenticated LDAP queries pull zimbraPreAuthKey, zimbraAuthTokenKey, zimbraTwoFactorAuthSecret. A dedicated Go implant (embedded module path zimbra-exfil/client-dump) reads /opt/zimbra/conf/localconfig.xml, extracts eight named service passwords, exports full tables from mailbox, mailbox_metadata, mobile_devices, out_of_office, and all zimbra.* tables, stages certs/keys/Postfix LDAP configs, ZIPs everything under /tmp/zimbra_dump_YYYYMMDD_HHMMSS/, and prepares transfer to a hex-encoded command-line argument destination. The crown-jewel read: zimbraAuthTokenKey = forge session tokens for ANY account without credentials; zimbraPreAuthKey = pre-authenticated login URLs for any user. Mail-server compromise here is identity compromise, not mailbox compromise.

Remote-access stack (one campaign): agent2.sh downloader → zimdown2 Go installer (SHA-256-verified fetch, WebSocket status channel, environment-selected paths/filenames) → zimclient2 agent: interactive shell, bidirectional file ops, SOCKS5 proxying, WebSocket/TLS/raw-TCP transports. Persistence breadth across the campaign: systemd, OpenRC, cron (incl. @reboot), shell startup files, authorized_keys, local-account creation.

Reverse shell shape: named pipe at /tmp/s feeding /bin/sh into openssl s_client, encrypted, utility-only C2.

Exfiltration attempt: mailbox-backup archive at /opt/zimbra/final.tar.gz, then AzCopy downloaded from hxxps://aka[.]ms/downloadazcopy-v10-linux invoked with an operator-supplied Azure Blob SAS URL to wsweb03[.]blob[.]core[.]windows[.]net/log/windows.log (transfer not confirmed). Read: legit Microsoft CDN + legit Microsoft transfer tool + a SAS URL = an egress event no proxy policy flags by reputation.

Defender actions

  1. Patch to 10.1.20+; if the zimbra-snmp package or SNMP notifications are present on any version, treat upgrade as urgent, and check whether the package was ever installed (removal is a valid interim control if SNMP notifications are unused).
  2. Hunt: swatchdog/snmptrap child anomalies; JSP files in Jetty/mailboxd paths with deleted-fragment neighbors; /etc/sudoers or sudoers.d entries for zimbra; pam.d/sudo modification history vs zmmailboxdmgr log-directory symlink checks; zimlog.service and any systemd unit whose mtime suspiciously matches neighbors; /tmp/s pipes; openssl s_client from zimbra; zmlocalconfig -s + ldapsearch pairs in shell history/logs (the dump sequence itself is the IoC); /tmp/zimbra_dump_* directories; zmprov + zimbra_identity SSH from non-automation sources; azcopy invocations with external SAS URLs.
  3. Post-incident assumption: if a Zimbra host ran a vulnerable version exposed to SMTP, the pre-auth and auth-token keys must be rotated, not just the host rebuilt, token forgery survives host reinstalls if keys persist.
  4. Correlate with on-wiki Zimbra history: CL-STA-1114 espionage page and the August KEV Zimbra row; this is now the third distinct Zimbra exploitation track on-wiki.

Sources