BeyondTrust Remote Support: patching closes the hole, not the exposure
A zero-day in BeyondTrust Remote Support was exploited from late August 2026 and disclosed on 15 October. The appliance brokers privileged credentials, so the remediation question is not whether the patch is applied but what left the box before it was.
What happened
BeyondTrust disclosed CVE-2026-102764 on 15 October 2026, a critical flaw in Remote Support that lets a remote attacker perform file operations on the appliance without authentication or user interaction. The vendor’s advisory index records it as BT26-05, CVSS 9.3 [2]. Privileged Remote Access shares the affected code; testing found no exploitable path there and a precautionary patch shipped anyway [1][7].
The timeline matters more than the severity score. Exploitation began on 29 August 2026 against one hosted instance and one self-hosted appliance. On 27 September a separate campaign against another vendor’s edge appliances became public and Mandiant sent BeyondTrust a non-specific alert. Threat hunting found the activity on 28 September, the hosted instance was contained 27 minutes later, and product security reproduced the entry vector that evening [1].
The patch dates diverge. BeyondTrust says it released the patch on 8 October, with SaaS instances patched automatically [1]. Google Threat Intelligence Group, SOCRadar and Cybersecurity Dive all give 9 October for SaaS tenants [3][4][6]. Release on the eighth and completion on the ninth is the likely reading, but the two dates appear as a single fact in different accounts.
CISA added the CVE to its Known Exploited Vulnerabilities catalogue on 16 October, with a remediation date of 19 October for United States federal civilian agencies, in the same batch as CVE-2025-61624, a path traversal in the FortiOS auto-script engine [5].
Three names for one defect
Each authority describes the flaw differently, and the differences are not cosmetic.
- BeyondTrust and the advisories built on it describe improper origin validation, CWE-290, in an internal messaging component that does not verify an incoming message came from a trusted source [6][8].
- Google Threat Intelligence Group and Mandiant describe the impact as unauthenticated arbitrary file upload leading to host-level control and access to stored credentials, in all supported versions up to 26.2.3 [3].
- The catalogue entry is titled authentication bypass by spoofing [5].
Each describes a different layer of the same bug, and a defender reading only one will scope their hunt wrongly. The vendor also notes that variant analysis found additional paths to the same weakness, and that the patch closes those too [8]. A detection built around one observed request pattern does not cover the class.
What the intrusion did, and why it changes the remediation
Google Threat Intelligence Group and Mandiant published their analysis the same day as the advisory, drawn from a Mandiant incident response engagement [3]. The sequence is short, entirely post-exploitation, and each step maps to a remediation step that patching does not perform.
- A web shell was placed in the web root about eleven minutes after the initial requests, written to return a not-found response unless a specific parameter was supplied, which defeats naive scanning [3].
- Root was reached with a set-user-ID copy of a shell under a system-sounding name, and persistence written into the appliance’s boot-time configuration, so passwordless administrative access is rebuilt on every restart. Reboot and configuration reset do not remove it [3][7][8].
- Credentials were pulled from the appliance’s own secrets store using the product’s own command-line utility, then appliance, web server, interpreter and authentication logs were deleted [3][7].
- Compromised administrator credentials raised session and team roles to administrator inside the product’s own console, after which support sessions and the file-transfer service reached internal endpoints and removed documents [3].
- A SOCKS5-over-WebSocket tunneller was deployed and the appliance’s reverse proxy reconfigured so one web path reached it. The backend pointed at localhost, hiding the listener from external scanning [3].
The population that has to act
Confirmed impact is narrow. BeyondTrust names two affected customers and says no adversary activity has been seen since August [1]. Kelli Vanderlee, senior analyst at Google Threat Intelligence Group, told Cybersecurity Dive that it is aware of only a small number of impacted organisations, while warning others may have been compromised before disclosure [4].
The exposed population is not. When BT26-02 was disclosed in February 2026, a pre-authentication command injection rated CVSS 9.9 in Remote Support 25.3.1 and earlier, the researcher who found it put roughly 11,000 instances on the internet, about 8,500 of them self-hosted [9]. SaaS customers are patched for them. The self-hosted majority is not, and appliances older than 24.3.1 must be upgraded to a supported release before the BT26-05 patch can be applied at all [6][7][8].
This is the fourth critical or high advisory in this product line in under two years. The vendor’s index records BT24-10 at CVSS 9.8 in December 2024, BT26-02 at 9.9 in February 2026, BT26-03 with four CVEs in July 2026, and now BT26-05 [2]. The Canadian Centre for Cyber Security issued its own advisory on the July set, covering both products at 25.3.2 and prior [10]. Treated individually these are patching events. Listed together they are a trend line.
What to do
Guidance across the vendor, the researchers and the independent advisories converges tightly, and patching is the first step rather than the whole job [1][3][8].
- Apply BT26-05 to every self-hosted Remote Support appliance, and to self-hosted Privileged Remote Access as a precaution. Upgrade first if it predates 24.3.1.
- Treat any appliance internet-reachable on a vulnerable version during the window as potentially compromised. Upgrading says nothing about prior access.
- Preserve evidence first: snapshot with memory, export disks, collect external proxy, web application firewall and firewall logs. Then rebuild rather than clean, because the persistence survives reset.
- Rotate every credential the appliance stored, injected or brokered, and only after patching or rebuilding: administrator accounts, API secrets, SAML keys, vault and integration credentials, anything entered during sessions. Rotating into a compromised appliance hands over the new secrets.
- Restrict the administrative and appliance interfaces to trusted networks, require federated sign-in with phishing-resistant multi-factor authentication, and review new registrations on privileged accounts.
- Forward logs off the appliance with multi-month retention. Local logs were deleted, so off-box telemetry is the only reliable record, and a forwarding gap bounded by one session is itself an indicator.
- Hunt behaviour: privileged sessions with no matching administrator or ticket, unrecognised scripts in web-accessible directories, service accounts spawning shells, role changes without a change record, and changes to the appliance’s proxy and privilege configuration [6][7].
The pattern worth carrying forward
Google Threat Intelligence Group assesses with high confidence that this actor will continue using edge-device zero-days, in a pattern of selective zero-day use followed by broad opportunistic exploitation once a flaw is public, citing its observations after the 2024 Ivanti Connect Secure disclosures as precedent [3]. Mandiant has previously linked the actor to Citrix NetScaler and Ivanti Connect Secure exploitation [4][7].
The deeper point is what class of box this was. Remote support and privileged access platforms exist to hold and broker other systems’ credentials. When one falls, the exposure is everything the appliance could reach, and that clock started before anyone knew the flaw existed. Unit 42’s 14 October analysis of a management-plane compromise makes the same point from a different product: configuration data for more than 500 managed firewalls was taken from one platform, and the researchers state plainly that patching it does not end the exposure of the credentials and certificates it held [11].
The offensive-security reading is unglamorous. The highest-value target in most estates is not the crown-jewel application. It is the thing with standing access to all of them, and it is usually treated as infrastructure rather than an asset with its own blast radius. Inventory those, restrict them at the network, ship their logs where they cannot be deleted, and decide in advance whether you would rebuild or clean. Deciding during an incident is deciding badly.
Sources
Every R3KONX article cites its primary material. 11 sources, in order of first citation. Links open the original publication.
- ARMORY CASTLE and BT26-05: What Security Teams Need to Know BeyondTrust · 2026-10-15
- Security Advisories index (BT26-05 and prior advisories) BeyondTrust · accessed 2026-10-17
- ARMORY CASTLE Exploits BeyondTrust Google Threat Intelligence Group and Mandiant · 2026-10-15
- China-nexus espionage actor targets BeyondTrust support tool in zero-day campaign Cybersecurity Dive · 2026-10-16
- U.S. CISA adds Fortinet Multiple Products, and BeyondTrust Remote Support flaws to its Known Exploited Vulnerabilities catalog Security Affairs · 2026-10-17
- CVE-2026-102764: BeyondTrust Zero-Day Exploited SOCRadar · 2026-10-16
- CVE-2026-102764: Critical BeyondTrust Remote Support Zero-Day Exploited by China-Linked Hackers SOC Prime · 2026-10-16
- Critical Vulnerability in BeyondTrust Remote Support Under Active Exploitation (CVE-2026-102764), BSL-A1231 Beazley Security Labs · 2026-10-15
- BeyondTrust Fixes Critical Pre-Auth RCE Flaw (BT26-02, CVE-2026-1731) The Hacker News · 2026-02-09
- BeyondTrust security advisory (AV26-660) Canadian Centre for Cyber Security · 2026-07-06
- FortiManager CVE-2024-47575 and the PEERSTITCH FortiGate implant Unit 42, Palo Alto Networks · 2026-10-14
Researched and written by the R3KONX analysis desk from the cited primary material. Methodological caveat: the vendor has withheld technical detail until self-hosted appliances are patched, so the flaw is described here only at the level the vendor, the researchers and the national advisories state; three of those sources characterise the same defect differently, and all three characterisations are published below rather than reconciled. CISA's own alert pages could not be retrieved from this network and are not cited. Attribution of the intrusion is reported as the assessment of the named researchers. Corrections to event@r3konx.asia.
