Patched Is Not the Same as Safe: Reading Version and Exposure for BeyondTrust Appliances

The BeyondTrust advisory for CVE-2026-102764 states the affected range plainly: Remote Support 26.2.3 and earlier, with a fix available. That sentence reads like a simple instruction, and for an appliance that nobody touched during the exploitation window it is one. For everything else, three different questions get collapsed into a single patch ticket.
Those questions are: which version is running, whether the appliance was reachable from the internet, and whether anything changed on it while the flaw was live. Only the first is answered by a patch.

Question one: version

Version is the easiest of the three and still the most error-prone in practice. Self-hosted installations require manual patching unless automatic updates are enabled, which means the fix arrives when an administrator schedules it, not when the vendor ships it. On an appliance installed by a regional team and administered from a different office, that gap can be months.
Version also has an edge problem in the public record: the advisory describes the affected range as 26.2.3 and earlier but does not enumerate every earlier build that is exposed. Owners should treat the vendor's statement as authoritative for their own build rather than inferring safety from a version number that merely looks newer.

Question two: reachability

Reachability decides whether the exploitation window applies at all. Mandiant's guidance is explicit that any appliance reachable from the internet on a vulnerable version during the window should be treated as potentially compromised until reviewed.
Whether an appliance is reachable is a fact an organisation can establish for itself, and external measurement gives a cross-check on the wider population. ZoomEye queries on 2026-10-23 returned these counts.
| Query executed 2026-10-23 (UTC) | Matching assets |
| --- | ---: |
| title="BeyondTrust" | 966 |
| title="BeyondTrust" && is_domain=true | 584 |
| app="BeyondTrust Privileged Remote Access" | 226 |
| title="BeyondTrust" && port="443" | 269 |
| title="BeyondTrust" && port="80" | 203 |
| title="BeyondTrust Remote Support" | 3 |
A word on the last row: three matches for the precise product name is not evidence that the product is rare. It is evidence that the fingerprint is thin. Anyone reading the 966 figure as a vulnerable count would be making the opposite mistake, because at least 584 of those matches are website assets rather than appliances.

Question three: integrity

This is the question a patch does not answer, and the incident shows why. The actor reached root on the appliance, read the local credential store through the native secrets utility, granted passwordless root to the site service account by editing configuration templates, and raised Jump Item and Team permissions to Administrator.
Two of those changes are configuration, not files. They survive a reboot, and Mandiant notes they survive a configuration reset. A patched appliance carrying them looks healthy from every version check in existence.
The practical consequence is that the response cannot stop at updating a package. Google's recommendation to redeploy appliances with confirmed or unknown compromise status from a clean, fully patched image exists because integrity has to be re-established, not restored.

Ordering the work

A sequence that keeps the three questions distinct:

1. Inventory appliances, including ones that no longer appear in project documentation. External queries help find the forgotten ones.
2. Establish reachability from the public internet for each, using firewall configuration rather than assumption.
3. Determine version and patch. Record the build number, not the ticket status.
4. For every appliance that was reachable and vulnerable during the window, treat integrity as unknown and rebuild from a clean image.
5. Rotate credentials in dependency order afterwards, and invalidate session material by redeploying Jump Clients, Jumpoints and the representative console.
6. Restrict /login and /appliance at the network level so that the next incident starts with a smaller population. The temptation is to compress steps three to five into "patched". That compression is where incidents get closed before they are understood.

What the record does not settle

No published figure says how many appliances were compromised, and no victims have been named. Whether credential harvesting succeeded in any specific case is unconfirmed, though access to the store is described. Attribution is characterised as suspected. The vulnerability's root cause is withheld by the vendor.
Those gaps argue for caution in both directions: not declaring an estate clean because every version number looks right, and not declaring it compromised because a query returned a large number. The measurement tells you where to look. The rebuild is what actually answers the question.

References

• BeyondTrust advisory BT26-05: https://www.beyondtrust.com/trust-center/security-advisories/bt26-05
• Google Threat Intelligence, Armory Castle exploits BeyondTrust: https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-beyondtrust
• BleepingComputer, BeyondTrust warns of critical Remote Support flaw exploited in attacks: https://www.bleepingcomputer.com/news/security/beyondtrust-warns-of-critical-remote-support-flaw-exploited-in-attacks/
• Infosecurity Magazine, China Espionage Actor Actively Targeting Critical BeyondTrust Flaw: https://www.infosecurity-magazine.com/news/china-actor-beyondtrust-flaw-google/
• ZoomEye search used above: https://www.zoomeye.ai/searchResult?q=dGl0bGU9IkJleW9uZFRydXN0Ig%3D%3D