The most useful detail in SOCRadar’s analysis of an exposed ransomware operation is not the artificial intelligence, even though the AI is what everyone is talking about. It is a password.
In one victim environment, researchers found that a single reused credential gave the attackers the VPN, the RMM, the backup platform, the security management console and the servers sitting behind all of them. <cite index=”0-1″>SOCRadar describes credential reuse as the main force multiplier of the operation.</cite> Not a zero-day. Not a novel technique. One password, entered in five places by whoever set the environment up, and never revisited.
The AI part is real, and we will get to it. But the honest reading is deflationary: the models did not invent any of this. They shortened the gap between one step and the next.
What Happened
SOCRadar’s Threat Research Unit found an open directory on a command-and-control server at 45.227.254[.]43, left exposed by operators who were careless with their own infrastructure. <cite index=”0-3″>Inside were the operators’ AI session databases, their knowledge base and the source code of a framework they call Phantom, described in their own documentation as an “Autonomous Red Team Kill Chain Framework”.</cite> <cite index=”0-4″>Configuration files, logs and credentials on that host pointed researchers to six more adversary-controlled servers.</cite>
<cite index=”0-6″>The operators are affiliates working with the INC and The Gentlemen ransomware brands.</cite> <cite index=”0-30″>SOCRadar analyzed 8,257 messages across 133 sessions in the exposed logs.</cite>
The operating model is the part worth understanding. <cite index=”0-28″>SOCRadar characterizes it as AI-executed and human-approved: the person directs and signs off, while the AI stack plans reconnaissance, picks techniques, generates commands, sorts through stolen data and writes the reports.</cite> <cite index=”0-31,0-32″>Each campaign opened with a master prompt written for that specific victim, and the model was instructed to wait for approval at every step.</cite>
Two details stand out for anyone drafting an AI policy. First, these were not underground models. <cite index=”0-33″>They were ordinary commercial LLMs accessed through OpenRouter, and the “WormGPT” branding found on the servers was only a persona prompt.</cite> Second, the guardrails worked, and then did not matter. <cite index=”0-33″>When a model declined a request, the operator did not go shopping for a more permissive one. He reframed the task as an “Authorized Red Team Assessment Request” and submitted it again.</cite>
None of this was cheap for them, incidentally, but it was not expensive either. <cite index=”0-35″>A check on October 5, 2026 found two paid OpenRouter accounts, with $23,920 in recorded spending.</cite>
Initial access came through the unglamorous edge of the network. <cite index=”0-45,0-46,0-47″>Operator records, which SOCRadar notes are not independently verified, claim 2,236 targets and 108 created administrator accounts through a FortiClient EMS authentication bypass (CVE-2026-35616), 427 harvested admin passwords from scans of 4,922 SimpleHelp servers (CVE-2024-57727), and more than 49 claimed breaches of small hosting providers through cPanel (CVE-2026-41940).</cite> <cite index=”0-46″>Stolen MSP technician credentials, SOCRadar notes, opened the door to every client those technicians supported.</cite>
<cite index=”0-73″>Of 62 named target organizations, 36 were in the United States, led by industrial, manufacturing and energy firms, then healthcare and pharmaceuticals.</cite> <cite index=”0-75″>Among the targets whose size could be classified, 27 had revenues of $250 million or less.</cite> This was not a hunt for the Fortune 100.
Why This Matters for Businesses
Strip out the AI and you are left with a very old problem wearing new shoes.
The five systems that one password opened are not five systems of equal value. The VPN is the front door. The RMM is the thing that can push software to every endpoint you own. The backup console is your recovery plan. The security management platform is how you would notice any of it happening. An attacker who reaches those last three has already won the argument before encryption starts.
That is why the encryptor behaves the way it does. <cite index=”0-62″>SOCRadar’s analysis of the INC Windows build describes it stopping backup and database services and deleting shadow copies through direct driver calls rather than the usual vssadmin commands.</cite> The tooling assumes it will have the access it needs to dismantle recovery. Credential reuse is what hands it that access.
The extortion economics have moved too. <cite index=”0-69″>The recovered records show a data-first playbook: estimate what the victim can pay from its own stolen financial documents, send masked proof samples, open with a quiet email to an owner or executive, then escalate through vishing calls, customers, partners, regulators and finally publication, on deadlines of 24 to 72 hours.</cite> <cite index=”1-20″>The Hacker News notes the operators specifically sized demands to what a victim could pay without pulling in a bank or an insurer.</cite> If your plan for a bad week is “we restore from backup and say nothing”, that plan has a gap in it.
And then there is the product list: FortiClient EMS, SimpleHelp, cPanel, VPN gateways. These are the systems a small IT team installs once, gets working, and stops thinking about. They are internet-facing by design, they hold administrative power, and they rarely appear on anyone’s asset inventory with a patch owner next to them.
What Organizations Should Consider
The fix for the central problem costs almost nothing. It is just tedious, which is why it does not get done. A few questions worth putting to your IT team or provider this month:
If our RMM credential leaked tonight, what else would it open? Ask for a straight answer, not a reassurance. Then ask the same question about the backup console and the security console.
Do the backup platform and the security console have their own identities? Separate accounts, unique passwords, MFA enforced, not reachable with the same domain admin that administers file servers. <cite index=”0-85″>SOCRadar’s own recommendations call for dedicated admin accounts, MFA and IP allowlists on MSP and RMM environments, and for requiring the same of any third-party provider.</cite>
Are our backups immutable or air-gapped? <cite index=”0-88″>That is explicitly on the recommendation list, alongside isolating and patching backup consoles.</cite> A backup an administrator can delete, is a backup an attacker can delete.
What management interfaces of ours are reachable from the internet right now? Not what should be: what is. Remote support tools, hosting control panels, VPN portals, EMS consoles. <cite index=”0-84″>The researchers suggest reviewing EMS specifically for administrator accounts nobody created and for altered VPN connection scripts.</cite>
Who patches the edge appliances, and how fast? If the answer is a name rather than a process, that is the finding.
What does our AI policy assume about refusals? This case is a useful argument for anyone who thinks model safety features are a security control. They are a speed bump for an attacker who knows how to phrase a request. Policy should govern what your people and your data can do with these tools, not rely on the vendor saying no.
One honest caveat on the numbers above: the target counts come from the attackers’ own notes, and attackers inflate. The structural lesson does not depend on whether the figure was 2,236 or a fraction of it.
How ISM Can Help
Identity & Access Management is where this story is actually won. Separating administrative identities, removing shared credentials from privileged platforms, and making sure the backup and security consoles cannot be reached with a password that also works on the VPN is unglamorous work with an outsized return.
Penetration Testing answers the question an internal review usually cannot: what an outsider can actually reach, and how far one working credential travels once they have it. Periodic testing matters more now precisely because AI compresses the time between discovery and exploitation. The window between “exposed” and “exploited” is shorter than your quarterly review cycle.
Cybersecurity & Risk Management covers the surrounding work: inventorying internet-facing systems, patch ownership for edge and remote-access tools, backup isolation, and tamper protection on endpoint defenses.
AI Readiness & Automation is the right home for the policy conversation, both for the tools your teams are already using and for a realistic view of how attackers use the same models.
Our own framing of this work is cyber-first managed IT: the security questions come before the convenience ones, because convenience is what put one password on five systems in the first place.
Talk With Our Team
If you want to know which of your administrative platforms share a credential, that is a conversation we have regularly and it rarely takes long to get a useful answer. Get in touch with us and we will start with your backup and remote management stack, since that is where this particular crew went first.
If you would rather take a look on your own before speaking to anyone, the Cyber Maturity Calculator gives you a sense of where your controls stand today.