CVE-2026-102764: is your BeyondTrust Remote Support exposed from the outside?

A remote access appliance is a credential vault with a map of everything it connects to. This one was exploited for 40 days before a patch existed. What happened, how to check your exposure without an exploit, and what BeyondTrust and Mandiant say to do.

Cutaway clay model of connected infrastructure components
One appliance, and a line to every system it is allowed to reach.
Quick answer

CVE-2026-102764 is a critical flaw (CVSS 4.0: 9.3) in BeyondTrust Remote Support that lets an unauthenticated attacker perform file operations on the appliance. It was exploited as a zero-day from August 29, 2026, and Mandiant attributes the activity to ARMORY CASTLE, a suspected China-nexus espionage actor. Self-hosted appliances up to 26.2.3 without the BT26-05 patch are exposed. Remote Support 26.3.1 and later include the fix.

40Days exploited before a patch existed
9.3CVSS 4.0, no authentication
1Day from CVE to CISA KEV

What is CVE-2026-102764?

A missing origin check inside BeyondTrust Remote Support that turns into file upload, root and the appliance's credential store. The CVE record says internal components "did not adequately validate the origin of the messages" they received, which "could allow an attacker to perform file operations on the appliance." Mandiant adds that it allows an unauthenticated remote attacker to upload arbitrary files, which can lead to host-level control and access to every credential the appliance stores or brokers.

BeyondTrust is withholding the technical details while self-hosted customers patch. That is the right call, and it is why this page has none either.

What Mandiant documented after initial access, at a high level:

Webshell

SAWFLY, an obfuscated PHP webshell that returns a 404 to anything without its secret parameter, so scanners walk past it.

Root and persistence

A SUID copy of the shell, plus a boot script that adds passwordless sudo and survives reboots and configuration resets.

Credentials

The appliance's own secrets tool used to dump the site and global vault, then its logs deleted.

Downstream access

Jump Item and Team permissions raised from "No Access" to "Administrator", then support sessions used to reach other systems and take documents.

Tunnel

BARECLAW, a SOCKS5-over-WebSocket proxy, published to the internet through the appliance's own reverse proxy.

When was it disclosed, and was it exploited before the patch?

Yes. Exploitation started 40 days before the patch and 47 days before public disclosure.

Date (2026)What happened
Aug 29The actor exploits the zero-day against one BeyondTrust-hosted instance and one self-hosted appliance.
Sep 27Another vendor's edge-appliance zero-day campaign goes public. BeyondTrust starts hunting its fleet for similar activity.
Sep 28BeyondTrust finds the hosted instance, contains it within 27 minutes, and reproduces the entry vector that evening.
Sep 30Mandiant shares findings from the self-hosted customer.
Oct 8BT26-05 patch released. SaaS instances are patched automatically.
Oct 15CVE published, advisory and coordinated disclosure.
Oct 16CISA adds it to KEV. Mandiant publishes its analysis.
Oct 19KEV remediation due date.

One line in BeyondTrust's write-up deserves more attention than it got. Its sensors fired in real time when the actor escalated privileges, and the automated triage in its third-party detection platform closed those alerts as false positives. The intrusion was found weeks later by a threat-intel-driven hunt. If a security vendor hunting its own fleet can lose a month that way, plan for it in yours.

The next phase is predictable. BeyondTrust and Mandiant both expect what this actor usually does once a vulnerability is public: move from a few chosen targets to every unpatched appliance it can reach. That is also the pattern in our KEV analysis: the 2026 CVEs that CISA confirmed as exploited on the day they were published are mostly network and security appliances. This one took a single day.

Does it matter for you? Three questions

If all three answers are yes, treat it as an incident until a forensic review says otherwise.

  1. Is it exposed? Remote Support is built to be reached by customers and technicians from anywhere, so the appliance itself is usually internet-facing. The sharper question is which parts answer from the internet: the customer portal, or also /login and /appliance.
  2. Is it exploited? Yes, by a state-linked espionage actor, with a CISA KEV listing. That puts it ahead of anything that is only vulnerable on paper.
  3. Is it exploitable here? Self-hosted Remote Support up to 26.2.3 without the BT26-05-RS patch: yes. Version 26.3.1 and later: fixed. SaaS: patched by BeyondTrust. Privileged Remote Access: no known exploitation path, but it shares the code, so patch it.

How do you check exposure from the outside without exploiting anything?

Find the appliance, see which interfaces answer from the internet, and confirm the patch from the inside. No exploit traffic needed.

  • Find every appliance. The hostname your customers use for support sessions is the obvious one. Also check DNS, certificate transparency logs and your attack surface tool for old or regional instances, and for Privileged Remote Access appliances.
  • Load /login and /appliance from outside. From a network outside your perimeter, open both paths in a browser. If either shows a sign-in page, it is reachable from the internet, and Mandiant's guidance is to restrict both by IP or network segment. That is a finding whatever the patch status.
  • Confirm the patch from a trusted network. In the /appliance interface, check that BT26-05-RS is applied or the appliance runs Remote Support 26.3.1 or later. Appliances older than 24.3.1 have to upgrade to a supported version first.
  • Write down the window. If an unpatched appliance answered from the internet between August 29 and the day you patched, that is your exposure window for the forensic review.

BeyondTrust's product had a security team, a detection stack and Mandiant working alongside it. The product has a million eyes on it. Whether your /appliance page answers a stranger on the internet is a question that usually has none.

How do you fix it, and what if you were already exposed?

Patch first. Then, if you were exposed, preserve evidence, rebuild rather than clean, and rotate credentials only after the appliance is trustworthy again.

  1. Patch. Apply BT26-05-RS from /appliance on supported versions, or upgrade Remote Support to 26.3.1 or later. Patch self-hosted PRA too.
  2. Preserve evidence. Snapshot virtual appliances with memory before changing anything, and collect syslog, reverse proxy, WAF and firewall logs for the whole window.
  3. Rebuild if in doubt. Redeploy from a clean, fully patched image, and restore only a configuration backup from before the window.
  4. Rotate in order. Appliance admin, /login users, API secrets, directory bind accounts, vault and PRA credentials, then integrations. Rotating before the rebuild hands the new secrets to whoever is still inside.
  5. Reduce exposure. Restrict /login and /appliance, allow outbound traffic only where the appliance needs it, require SAML SSO with phishing-resistant MFA, and set the default Jump Item role to No Access.
  6. Get the logs off the box. This actor deletes local logs. Forward syslog to your SIEM and keep months of it.

For hunting, Mandiant and BeyondTrust point at behavior over indicators: role changes on Jump Items or Teams, admin logins from residential or commercial VPN addresses, unknown .php files in the web root, WebSocket upgrades to unfamiliar paths at the proxy, and gaps in log forwarding. Mandiant's post has the file hashes and YARA rules.

A remote access appliance is a credential store. Compromise it and the attacker inherits every session it brokers.

An auto-closed alert is still an alert. Keep what your triage layer dismisses in scope for hunting.

Admin interfaces don't belong on the internet. Restricting /login and /appliance costs an afternoon.

FAQ

Was CVE-2026-102764 exploited as a zero-day?

Yes. BeyondTrust says it was exploited on August 29 against one hosted instance and one self-hosted appliance. The patch shipped on October 8 and the CVE was disclosed on October 15.

Are SaaS customers affected?

BeyondTrust patched all Remote Support and PRA SaaS instances automatically. Self-hosted appliances need the patch applied by the customer or an upgrade to Remote Support 26.3.1 or later.

Is Privileged Remote Access vulnerable?

BeyondTrust found no way to exploit it in PRA, but PRA shares the affected code, so the patch was released for PRA too.

Does the patch remove an existing compromise?

No. The persistence Mandiant found survives reboots and configuration resets. Treat any appliance that was reachable and unpatched during the window as potentially compromised.

What should be reachable from the internet?

Only what technicians and customers need: the representative console, the customer portal and Jump Client traffic. Not /login, not /appliance.

Ron Kagansky is the founder and CEO of Cylent, an AI-native offensive security company running continuous external red teaming. Every path our agents find is validated by an operator before it reaches a report.

Sources

  1. BeyondTrust, ARMORY CASTLE and BT26-05: What Security Teams Need to Know (October 15, 2026, updated October 16, 2026)
  2. BeyondTrust, Security Advisory BT26-05 (updated November 4, 2026)
  3. Mandiant / Google Threat Intelligence Group, ARMORY CASTLE Exploits BeyondTrust (October 16, 2026)
  4. CISA, Known Exploited Vulnerabilities Catalog, catalog version 2026.10.30
  5. CVE Program, CVE-2026-102764; NIST NVD, CVE-2026-102764
  6. Rapid7, CVE-2026-102764: Critical BeyondTrust Remote Support Authentication Bypass Exploited in the Wild (October 19, 2026)
  7. Cylent, Time to exploit in 2026 (CISA KEV analysis, updated monthly)

Know which admin pages answer the internet.

Our agents map what answers from the outside the way an attacker would, and an operator validates every path before it reaches you.

Get a POC