Cut-paper illustration of a server appliance with a peeled-back layer revealing red warning markers inside, beside a fresh clean appliance

Friday afternoon, CISA added CVE-2026-102764 to its Known Exploited Vulnerabilities catalog with a remediation deadline of October 19: three days, forensic triage flagged. If your BeyondTrust Remote Support appliance was reachable from the internet any time since late August, this is not a routine patch ticket. The patch exists. It is necessary. It is also not enough.

What landed

On October 15, BeyondTrust published advisory BT26-05 for a critical origin-validation flaw (CWE-290) in an internal messaging component of Remote Support, formerly Bomgar. The headline facts:

Fact Detail
CVE CVE-2026-102764, CVSS 9.3 critical
Effect Unauthenticated remote attacker performs file operations on the appliance
Affected Remote Support 26.2.3 and earlier
Fix BT26-05-RS patch for 24.3.1 through 26.2.3; SaaS patched October 9
Older branches Below 24.3.1, upgrade first, then patch
Zero-day window Exploited in the wild starting August 29, 2026
Attacker ARMORY CASTLE (formerly UNC5221), suspected PRC-nexus espionage actor

Google's Threat Intelligence Group has watched this actor hit edge appliances repeatedly and warns to expect widespread opportunistic exploitation now that the flaw is public, the same pattern it ran after the Ivanti disclosures in 2024.

Why this one is nastier than a normal patch

The exploit gives the attacker unauthenticated file upload. What Mandiant documented next is what turns this into a rebuild case rather than a patch case. Working through a real compromised appliance, they found a professional, appliance-native intrusion that used the product's own plumbing against it:

  1. A staging POST lands, and minutes later the SAWFLY webshell (mangoapi.php) appears at the web root. It answers only when a specific obfuscated parameter is present and returns 404 to everything else, staying quiet for scanners.
  2. The actor copies /bin/sh to /usr/bin/dbus-authd, makes it root-owned, and sets the SUID bit. Plausible filename, persistent root backdoor.
  3. The appliance's own secrets vault is probed with the proprietary /ns/sbin/bt-secrets binary to dump passwords, API keys, and endpoint credentials, then the bt-secrets logs are deleted.
  4. Persistence goes into the configuration template layer: a boot script at /home/bomgar/sbin/post_boot.d/zz-root-persist writes a passwordless-root sudoers entry into /ns/config/sudoers.template and forces the backend to rebuild the live sudoers file. This survives reboots and configuration resets.
  5. The actor logs into the admin portal, raises a compromised account's Jump Item and Team roles from No Access to Administrator, launches support sessions to internal endpoints, and uses the native file transfer to pull documents off user desktops.
  6. A custom SOCKS5-over-WebSocket tunneler (BARECLAW, dropped at /tmp/ingredi_ws) is wired into the appliance's HAProxy config so the path /uidchanarge routes to a localhost backend on port 19800. The listener binds to loopback only, so external scans see nothing.

The durable idea: appliances are black boxes, so rebuild is the remediation

Mandiant's guidance contains one sentence that should set the plan for every operator: customers cannot audit the appliance host file system, and host-level persistence survives reboot and configuration reset. This is not a Windows box where you can boot to known-good media and diff the file system. The persistence here lives in the configuration template layer, so a "factory reset" from the UI replays the attacker's sudoers entry right back into place. Patching closes the vulnerability. It does not evict anyone.

That gives a clean decision rule for any edge appliance incident: if the vendor gives you no forensic audit path, and the attacker demonstrated persistence that survives your reset procedures, the remediation is rebuild-from-image, not hunt-and-clean. Hunting still matters, but it determines scope, not the recovery method.

The playbook

Mandiant recommends a phased, patch-first approach. Work it in this order.

1. Inventory. List every Remote Support appliance you touch: self-hosted versus SaaS, version, and whether it was internet-reachable on a vulnerable version between late August and the patch.

2. Patch first. Self-hosted on 24.3.1 or later: apply BT26-05-RS from the /appliance interface. Below 24.3.1: upgrade, then patch. Patch not visible in /appliance? Open a support case citing BT26-05. SaaS customers: BeyondTrust applied it October 9, but the credential and configuration steps below still apply to you.

3. Triage in parallel. While patching proceeds, hunt for the published indicators:

  • File artifacts: mangoapi.php under the web root, /tmp/ingredi_ws, /usr/bin/dbus-authd with the SUID bit set, /home/bomgar/sbin/post_boot.d/zz-root-persist
  • Config changes: unexpected entries in /ns/config/sudoers.template, ACL or backend additions in haproxy.conf for paths you did not create
  • In /login: user_changed events raising a Jump Item or Team role, especially No Access to Administrator; user_added, api_account, or security_provider events with no change ticket; changes to outbound event or syslog destinations
  • Authentication and sessions: admin logins from residential broadband or commercial VPN ranges, logins outside business hours, file-transfer events, sessions started by accounts whose roles changed during the window
  • Anti-forensics: gaps in syslog forwarding, especially an interruption that starts and ends inside one session

4. If you cannot clear it, isolate at the perimeter. Block inbound traffic to the appliance's public address at your firewall, not in /appliance, so the configuration stays intact for review. First: snapshot the VM with memory and export the disk; for hardware appliances or SaaS tenants, request forensic preservation through BeyondTrust Support. Pull syslog, reverse-proxy/WAF, and perimeter firewall logs for the whole window, because on-appliance logs may already be gone. End active sessions and disable, do not delete, Jump Clients, Jumpoints, and API accounts so the audit trail survives.

5. Rebuild, don't clean. Redeploy from a clean, fully patched image. Restore only a configuration backup dated before the exposure window, issue a new TLS certificate, and redeploy Jump Clients, Jumpoints, and the representative console to invalidate existing tokens.

6. Rotate credentials, in dependency order, and only now. Rotating before the appliance is patched or rebuilt hands fresh credentials to whoever still owns it. Order: /appliance admin credentials, /login administrators and local users, API client secrets, security provider bind accounts and SAML signing keys, Remote Support Vault and integrated Password Safe credentials at the target systems, integration credentials (SMTP, syslog, ITSM connectors), and any credential typed into a support session during the window.

7. Harden the control plane while you are in there. Restrict /login and /appliance to trusted networks so only console, customer portal, and Jump Client traffic is internet-facing. Put a reverse proxy or WAF with full URI logging in front of the appliance. Require SAML SSO with phishing-resistant MFA for /login users, and set the default Jump Item role to No Access in every group policy. Forward appliance syslog to a SIEM over TLS with multi-month retention, and schedule exports of the /login session, team activity, and user audit reports.

Three traps worth naming. Do not rotate credentials before the appliance is patched or rebuilt; you are just re-arming the attacker's access. Do not delete suspect accounts during triage; disable them so the audit entries survive. Do not isolate by editing /appliance on a potentially compromised box; use your own firewall at the perimeter, where the attacker cannot reach the rule.

Sources

The patch is the easy part. The rebuild is the actual fix. Plan accordingly.