BeyondTrust Remote Support zero-day: patch it, then check who got in
Attackers used CVE-2026-102764 against BeyondTrust Remote Support from August, before a patch existed. Here's who needs BT26-05, how to tell if someone got in, and the order to rebuild and rotate.
part 4 of 7 · Platform releases, explained from scratchthe ideaPatch the appliance, check for intruders, rebuild, then rotate.
after reading, you can- Tell which BeyondTrust deployments need BT26-05 today and which ones BeyondTrust already patched
- Spot the account changes and log gaps that show someone got in before the patch
- Plan the snapshot, rebuild and credential rotation in the order that doesn't tip off an attacker
Attackers broke into BeyondTrust Remote Support in late August, through a flaw the vendor only patched in October. If you run Remote Support on your own appliance, patch it today and then check whether someone got in first. Patching won't remove anything an attacker already planted, so most of this post is about the second part.
Attackers got in through the remote tool your helpdesk uses. Someone has to check it today.
Remote Support (it used to be called Bomgar) is what many IT teams use to take over a user's screen and fix a problem. To do that job it keeps admin passwords and can open sessions to servers and laptops all over the company. CVE-2026-102764 let someone on the internet, with no login, put files on the appliance itself. In the intrusion Google's Mandiant investigated, that was enough to take root control of the box and go after the stored passwords. BeyondTrust patched its own hosted service in early October. If you run the appliance yourself, the fix is a patch you apply, and you still have to check whether someone got in before you did.
Picture the cabinet where a building's caretaker keeps every master key. It has a mail slot, and the slot didn't check who was posting things through it. Someone slipped in a tool that opened the cabinet from the inside. Fixing the slot stops the next visitor. It doesn't tell you whether keys were copied, which is why you also change the locks.
- Teams running Remote Support on-prem
- Helpdesk and IT support leads
- Security and incident response
- Owners whose IT provider uses it
- Self-hosted PRA customers
Why it matters to you
A remote support appliance is a very good place for an attacker to stand. It's reachable from the internet by design, it holds admin passwords, and it already has permission to open sessions to machines across the company. Anyone who takes control of it gets a ready-made way in to everything it can reach, and the traffic looks like normal support work.
The attack needs no password and no click from anyone. CISA's own assessment in the CVE record marks exploitation as active, the attack as automatable and the technical impact as total. It went into CISA's catalog of known exploited vulnerabilities on October 16 with a due date of October 19 for federal agencies, and with a flag that reads forensic triagedefinitionA first check of a system for signs of compromise, done before it's cleaned or rebuilt. CISA now requires it for some KEV entries, this one included. required. Under CISA's directive BOD 26-04, that means agencies have to check the appliance for signs of compromise as well as patch it. CISA's directive binds only federal agencies. It's still the right bar for anyone else running one of these boxes.
So far the confirmed victims are few. BeyondTrust says it knows of two compromises, both dating to August 29: one instance it hosted and one appliance a customer ran. Mandiant attributes the work to ARMORY CASTLE, which it describes as a suspected China-nexus espionage group, formerly tracked as UNC5221 (CrowdStrike calls it WARP PANDA). Mandiant also warns that wider exploitation attempts may follow now that the flaw is public. No public proof-of-concept had turned up as of October 16, and BeyondTrust is holding back the technical details while customers patch.
-
01 A request from the internet
The attacker sends web requests to the appliance. No account, no password and nobody clicking anything.
-
02 The appliance trusts the message
Internal components don't check where a message really came from, so the attacker can make it write files it should refuse.
-
03 A web shell appears
About 11 minutes after the first request, a small PHP backdoor Mandiant calls SAWFLY sat in the appliance's web folder.
-
04 Root, and it survives reboots
A copied shell with root rights, plus a startup script that adds a no-password sudo rule, keep the attacker in after restarts.
-
05 Tracks covered
Logs on the box were deleted and a hidden tunnel was routed through the appliance's own proxy, so it looked like normal traffic.
For engineers, Mandiant's report has the detail. The first requests went to an odd path, /app_84304470%2Ephp. The SAWFLY shell (mangoapi.php) answers with a 404 unless a secret parameter is present, which keeps scanners from spotting it. The attacker copied /bin/sh to /usr/bin/dbus-authd with the SUIDdefinitionA Linux file permission that makes a program run with its owner's rights. A copy of the shell owned by root with SUID set gives anyone who runs it root. bit set, dropped zz-root-persist into /home/bomgar/sbin/post_boot.d, and edited /ns/config/haproxy.conf so a WebSocket path, /uidchanarge, forwarded to a tunnel called BARECLAW on local port 19800. Then they ran /ns/sbin/bt-secrets against the vault and deleted bt-secrets.log along with more than a dozen logs.
The part that should worry an identity team comes after that. With a stolen administrator login, the attacker changed an account's Jump Item and Team permissions from No Access to Administrator. In Mandiant's words, that gave "permanent, unrestricted authorization to launch remote sessions to downstream systems." They used support sessions to ask for full file system access and pulled software inventories out. Nothing about that looks strange to a support tool, because opening sessions and moving files is its whole job.
What BeyondTrust fixed, and what's still on you
The fix is a patch called BT26-05-RS. BeyondTrust's advisory lists every Remote Support version up to and including 26.2.3 as affected. It applied the patch to all of its hosted Remote Support instances by October 9, so SaaS customers don't have to touch the appliance. Self-hosted appliances on 24.3.1 or later take the patch from the /appliance admin interface. Anything older has to be upgraded to a supported version first, and then patched.
What changed with BT26-05
BT26-05-RS closes the flaw on self-hosted versions 24.3.1 to 26.2.3. Apply it in /appliance, or confirm automatic updates already did.
BeyondTrust patched every hosted instance by October 9. You still need to review credentials and configuration, because the patch doesn't undo earlier access.
No direct patch. Upgrade to a supported version, then apply BT26-05-RS. Mandiant notes the upgrade closes the hole but doesn't clean up a past compromise.
PRA shares the affected code. BeyondTrust says it found no way to exploit it there, but it patched PRA too and asks self-hosted customers to apply it.
The patch only shuts the door. It doesn't tell you whether anyone came through it while it was open, and the attacker in Mandiant's case built persistence that survives reboots and configuration resets. Mandiant's advice is blunt: treat any internet-reachable appliance that ran a vulnerable version during the exploitation window as possibly compromised until someone has reviewed it.
The review is harder than it sounds. You can't audit the appliance's host file system yourself, so the file-level checks Mandiant lists mean working with BeyondTrust Support. If the patch doesn't show in /appliance, open a support case citing BT26-05. And the attacker deleted logs on the box, so anything you didn't already forward to a SIEM may be gone.
| Deployment | Fix | What you still do |
|---|---|---|
| Remote Support SaaS (hosted by BeyondTrust) | Patched by BeyondTrust by Oct 9, 2026 | Review admin accounts, permissions and stored credentials. Rotate where in doubt |
| Self-hosted Remote Support 24.3.1 to 26.2.3 | Apply BT26-05-RS from /appliance | Patch today, then hunt for the signs below and plan a rebuild if anything turns up |
| Self-hosted Remote Support older than 24.3.1 | Upgrade to a supported version, then patch | Treat as possibly compromised. The upgrade doesn't remove earlier persistence |
| Self-hosted Privileged Remote Access | Patch released, confirm the right one with Support | Apply it. BeyondTrust found no exploit path, but the code is shared |
How it unfolded
- Aug 29, 2026First known exploitation
One BeyondTrust-hosted instance and one self-hosted appliance.
- Sep 28, 2026BeyondTrust finds the hosted intrusion
- Sep 30, 2026Mandiant shares self-hosted findings
- Oct 8, 2026Patch released
- Oct 15, 2026Public disclosure
- Oct 16, 2026Added to CISA KEV
Forensic triage required for federal agencies.
- Oct 19, 2026Federal deadline
Agencies had three days to patch and check.
One detail from BeyondTrust's own account is worth reading twice. The company says its sensors fired in real time when the attacker escalated privileges, and the automatic triage in its third-party detection platform closed those alerts as false positives. If a security vendor's own alerts got closed by automation, assume yours could be too. Look again at anything your tooling auto-closed on these appliances since late August.
What to do now
What to do now, by who you are
- Ask IT or your provider one question: do we use BeyondTrust Remote Support or PRA, and is it hosted by BeyondTrust or run by us?
- If it's self-hosted, ask for written confirmation that BT26-05 was applied, and the date.
- Ask whether anyone checked the appliance for signs of a break-in. Patching alone doesn't answer that.
- If you have cyber insurance, read the notice clause now. Some policies want to hear about a suspected incident early.
- List every Remote Support and PRA appliance, including test boxes and B Series hardware nobody remembers.
- Take a snapshot with memory before changing anything, then apply BT26-05-RS in /appliance.
- Upgrade anything older than 24.3.1 first. Plan that as an emergency change, not the next quarterly window.
- Limit /login and /appliance to known admin networks, and cut outbound traffic to what the appliance needs.
- Hunt from August onward: admin logins from home broadband or VPN ranges, and logins outside support hours.
- Look for user_changed events raising a Jump Item or Team role, and new accounts or API accounts nobody ticketed.
- Check syslog for forwarding gaps that start and end around a single session. That pattern points to log deletion.
- Reopen anything your tools auto-closed on these appliances since August. BeyondTrust's own alerts got closed that way.
- Write down every password the appliance keeps or injects, and every machine its Jump Items can open. You'll need both lists later.
- Rotate those credentials only once the appliance is patched or rebuilt, in dependency order.
- Set the default Jump Item role to No Access in every group policy, then grant access back on purpose.
If you only have time for one check today, make it the account changes. Mandiant's case turned on a stolen admin raising permissions from No Access to Administrator, and that leaves a trail in the appliance's own events even when files on disk are gone. Pull the forwarded events for user_changed, user_added, api_account and security_provider, and match each one to a change ticket.
# example output. Field names differ by SIEM, so rename them to match yours # 1. Requests to PHP paths support staff never use (reverse proxy or WAF logs) uri ends_with ".php" and uri not in known_paths # 2. WebSocket connections on /uidchanarge (used here) or /wsproxy (the tunnel's default) uri in ("/uidchanarge", "/wsproxy") or header.upgrade = "websocket" and src not in support_ranges # 3. Gaps where the appliance stopped sending syslog. Someone may have deleted logs gap(host = rs-appliance, min = 10m)
The order that works for a self-hosted appliance
- 01Preserve first
Snapshot the virtual appliance with memory, or ask BeyondTrust Support how to capture a physical one. Export the syslog you already forward. A reboot or patch can wipe what an investigator needs.
- 02Cut exposure
Restrict /login and /appliance to admin networks at the firewall and limit outbound traffic. Censys points out this limits reach while you work. It doesn't fix the flaw.
- 03Patch
Apply BT26-05-RS from /appliance on 24.3.1 to 26.2.3, or upgrade older versions and then patch. If the patch isn't listed, open a support case that cites BT26-05.
- 04Hunt
Work through the signs in the roles block above, plus file indicators such as mangoapi.php, ingredi_ws, /usr/bin/dbus-authd and the zz-root-persist script. File-level checks need BeyondTrust Support.
- 05Decide: clean or rebuild
If anything turns up, or you can't rule it out because logs are missing, rebuild from a clean, fully patched image. Mandiant's reasoning is that host-level persistence survives reboots and configuration resets.
- 06Rotate, last
Rotate every credential the appliance stored or injected, after the patch or rebuild. Rotate too early and an attacker still on the box just reads the new ones.
Harden so the next one hurts less
Remote access appliances will keep getting bugs. Remote Support has had three serious advisories this year. Aim for the next one to cost you a patch night and a quiet log review, and nothing more. Most of the hardening below comes from Mandiant's recommendations for this case. Each item cuts off something the attacker in this intrusion actually used.
Before: break in once and you can reach everything
The appliance faces the internet, stores admin passwords and can open sessions anywhere, and its logs live only on the box.
After: patched, fenced and watched
Admin pages are reachable only from admin networks, credentials come from a vault at session time, and logs leave the box as they're written.
- /login and /appliance open to the internet
- Several local admin accounts with passwords
- Admin passwords sit in the appliance's own vault, so a dump gets them all
- New groups get access to Jump Items by default
- Logs kept on the appliance
- Patched in the next quarterly window
- Nobody outside the admin network can even load the admin pages
- SAML SSO with phishing-resistant MFA, one vaulted break-glass account
- Passwords come from Password Safe or another vault at session time and never stay on the box
- New groups start with No Access to Jump Items, so someone has to grant it
- Syslog forwarded to a SIEM over TLS, kept for months
- An emergency patch path for edge appliances
Two of these do most of the work. Getting stored credentials off the appliance means a vault dump gets the attacker very little, and the injection model means technicians never see the passwords anyway. Setting the default Jump Item role to No Access means a raised permission stands out as a deliberate change in the logs. Filter egress too: Mandiant suggests allowing only BeyondTrust's own endpoints and your internal syslog and directory servers, so a tunnel like BARECLAW has nowhere to call.
Keep a written list of what the appliance can reach. When the next advisory comes, that list tells you on day one which credentials to rotate and which servers to look at, instead of working it out during the incident. If you already run CyberArk or HashiCorp Vault for other secrets, the same rule applies: the remote access tool should borrow credentials at session time, never keep its own copies.
If you've been breached
If you find signs someone got in
Work through these in order. The two most common mistakes are rebooting before you capture evidence and rotating passwords while the attacker is still on the box.
- 01Containfirst hour
- Snapshot the appliance with memory before any change or reboot, and save all forwarded logs.
- Block its public address at the firewall, and cut outbound traffic down to what you need for the investigation.
- Disable active sessions and access paths. Disable them, don't delete them, so the evidence stays.
- Open a Sev 1 case with BeyondTrust Support that cites BT26-05. You'll need them for host-level checks.
- 02Investigatefirst days
- Pull syslog, reverse proxy or WAF logs and perimeter firewall logs for the whole window since late August.
- Trace every session launched by accounts whose roles changed, and every file transfer, to the machines it touched.
- Check those downstream servers and laptops for new accounts, scheduled tasks or tools nobody installed.
- Write down what the appliance could reach and which credentials it held. That list is your worst case.
- 03Rebuilddays 1 to 3
- Rebuild the appliance from a clean, fully patched image. Don't clean it in place: host persistence survives resets.
- Restore configuration selectively and recheck every account, API account, security provider and Jump Item role.
- 04Rotateafter the rebuild
- Rotate every credential the appliance stored or injected, now that the attacker can't read the new ones.
- Go in dependency order, as Mandiant advises: accounts that can reset other accounts come first.
- Revoke API keys and tokens the appliance used, including any for integrations and SIEM forwarding.
- 05Report and learnfirst week
- Bring in legal early. Depending on your sector and what the attacker reached, you might owe regulators or customers a notice, and the clock usually starts when you find out.
- Call your insurer early. Most policies have a notice window, and if you miss it the claim can get harder.
- Fix the gaps that let it happen: admin UI exposure, logs on the box, stored credentials, slow patching.
A note on rotation order, because it's where incident teams usually slip. The instinct is to change every password the moment you suspect a breach. Mandiant warns against it here: until the appliance is patched or rebuilt, an attacker who still controls it can read the new credentials as easily as the old ones. Contain, rebuild, then rotate.
What this could look like: three typical setups
What this could look like: three typical setups
- the setup
- A manufacturer's helpdesk uses BeyondTrust's hosted Remote Support. Technicians sign in with local accounts, and the vault holds a few domain admin passwords.
- what happens
- BeyondTrust patched the service by October 9, so the hole is closed. Nobody has checked whether an admin account or vault entry was touched before then.
- what to do
- Review admin accounts, API accounts and permission changes since August. Move the vault entries to injection from a real vault, switch admins to SSO with MFA, and rotate anything that looks off.
- the setup
- A regional hospital runs Remote Support 25.x as a virtual appliance at the edge, with the admin page open to the internet. Logs stay on the box and patches go in quarterly.
- what happens
- The appliance ran a vulnerable version through September with no off-box logs. If an attacker deleted local logs, the team can't prove nobody got in.
- what to do
- Snapshot, then patch today. Open a support case for host checks. With no logs to clear it, plan a clean rebuild, then rotate every stored credential and start syslog forwarding.
- the setup
- A bank runs Remote Support and PRA on-prem. Admin pages only answer from the admin network, credentials come from Password Safe, and syslog goes to the SIEM.
- what happens
- Requests from the internet never reach the admin pages, so a probe stops at the firewall. The SIEM shows no role changes, and the vault holds nothing an attacker could dump.
- what to do
- Apply BT26-05 to both products anyway, run the hunts against the SIEM back to August, and write down the result for auditors. Keep the emergency patch path.
These three are made up, and none of them is a real incident or a Cohash client. They all get the same patch from BeyondTrust. What decides how bad it gets is how each one was set up before August, mostly who could reach the admin page and whether logs left the box.
How Cohash helps
We work on BeyondTrust estates, mostly Privileged Remote Access, Password Safe and vendor access through jump points. On this advisory we'd start by finding every appliance, then plan the patch and the checks, and only rotate credentials once it's safe. Later we'd move stored passwords to injection and lock the admin pages down. The BeyondTrust page is at /tools/beyondtrust, engineers are at /tools/beyondtrust#hire, and you can book a review through /assessment.
What to take from this
- 01
Self-hosted Remote Support up to 26.2.3 needs BT26-05-RS today. Older than 24.3.1, upgrade first and then patch.
- 02
Patching doesn't remove an attacker who's already in. Treat exposed appliances as possibly compromised until they've been reviewed.
- 03
Snapshot before you patch or reboot, and rebuild from a clean image if you can't rule a compromise out.
- 04
Rotate stored and injected credentials after the patch or rebuild, never before.
- 05
If you're on the hosted service, you already have the patch. Someone still has to look over accounts, permissions and stored credentials.
- 06
Fence the admin pages, keep credentials off the box and get logs off it too, so the next zero-day costs a patch night.
- How Cohash delivers BeyondTrust: scope, roles and architecture
- How Cohash delivers CyberArk: scope, roles and architecture
- How Cohash delivers HashiCorp Vault: scope, roles and architecture
- Hire vetted BeyondTrust engineers
- Privileged access management
- Book a 30-minute working session with a senior engineer
01Am I affected by CVE-2026-102764?
If you run BeyondTrust Remote Support on your own appliance at version 26.2.3 or earlier without the BT26-05-RS patch, yes. BeyondTrust patched its hosted service by October 9. Privileged Remote Access shares the code, and BeyondTrust found no way to exploit it there, but it still asks self-hosted PRA customers to patch.
02Our IT provider manages BeyondTrust for us. Do we need to do anything?
Ask two things in writing: whether BT26-05 is applied and when, and whether anyone checked the appliance for signs of compromise. The second matters as much as the first, because the patch doesn't remove an attacker who got in earlier.
03We're on Remote Support 23.x. Can we just apply the patch?
No. The BT26-05-RS patch covers 24.3.1 to 26.2.3. Older appliances have to be upgraded to a supported version first, then patched. Treat the appliance as possibly compromised while you do it, since the upgrade won't clean up earlier access.
04Is there a public exploit for CVE-2026-102764?
None had been found as of October 16, and BeyondTrust is holding back technical details while customers patch. The flaw was used as a zero-day from late August, though, and Mandiant warns wider attempts are possible now that it's public. Don't wait for a public exploit before you patch.
05We've already patched. Could someone have got in before that, and how would we know?
Start with forwarded logs: admin logins from home broadband or VPN ranges, role changes on Jump Items or Teams, new accounts without a ticket, and gaps in syslog. File-level checks on the appliance itself need BeyondTrust Support, so open a case citing BT26-05.
06Should we rotate passwords right now?
Only after the appliance is patched or rebuilt. If an attacker still controls it, they can read the new passwords too. Once it's clean, rotate everything it stored or injected, in dependency order, so accounts that can reset others go first.
07CISA's KEV entry says forensic triage required. What does that mean for us?
BOD 26-04 is CISA's directive for federal agencies. For an entry like this one they had three days to patch and had to check the appliance for signs of compromise too. Nobody else is bound by it, but it's a sensible bar if you run an exposed appliance.
Three questions, about a minute
Pick an answer and we'll tell you if it's right, and why it is or isn't. The score stays in this browser.
-
Q1Your self-hosted Remote Support appliance runs 25.2 and faced the internet all autumn. You've just applied BT26-05-RS. What's next?
show the answer
C. Hunt for signs of compromise, and rebuild from a clean image if you can't rule it out The patch stops new break-ins but doesn't remove persistence planted earlier. Exposed appliances count as possibly compromised until reviewed, and 24.3.1 is inside the affected range.
-
Q2Your appliance shows signs someone got in. At what point do you rotate the passwords it stored and injected?
show the answer
A. After the appliance is patched or rebuilt An attacker still on the box can read new credentials as easily as old ones. Contain, rebuild, then rotate in dependency order, treating every stored or injected credential as exposed.
-
Q3We're on the hosted service and nobody here runs the appliance. Which of these is true for us?
show the answer
D. BeyondTrust patched it, but you should still review accounts, permissions and stored credentials BeyondTrust patched every hosted instance by October 9, and one of the two known victims was a hosted instance. Reviewing accounts and credentials is still on the customer.