Your SOC playbook quietly assumes the attacker gets tired, makes typos, and logs off for the night. An affiliate that delegates the boring steps of a ransomware affiliates AI agents intrusions workflow does none of those things. So the gap between first alert and encryption is the number to re-measure. Two research teams recently got an unusual look at that workflow after attackers left their own servers open on the public Internet.
Who Found the Exposed Servers
In late September 2026, the SOCRadar Threat Research Unit (STRU) found a publicly accessible command-and-control server at 45.227.254[.]43. Configuration files, logs, and credentials on that host led them to six more adversary servers, most on AS267784 (Flyservers S.A.). The Atos Threat Research Center (TRC), in a separate report by Piotr Mazurkiewicz, recovered material from an Internet-accessible host tied to the same INC ransomware ecosystem. Both teams describe affiliates of INC, also tracked as GOLD IONIC or MITRE G1032, with some cases branded as The Gentlemen. STRU rates confidence in a formal dual-brand alliance as low, and confidence that operators of both groups used a shared environment as moderate-to-high.
What they found was not a malware dump. STRU recovered AI session databases, a central knowledge base, and the source of Phantom, which the operators called an “Autonomous Red Team Kill Chain Framework.” Atos TRC recovered a 1,882-record knowledge export, about 846 MB of tooling and victim data across 444 files, and HTTP logs showing successful downloads from the exposed service. STRU says the operators logged 8,257 messages across 133 OpenCode sessions and that one OpenRouter account showed about $23,920 in recorded spending. Those figures are vendor findings from the recovered logs, not independent measurements. STRU also graded 62 named organizations by depth of activity, 36 of them in the United States, and notes that many of the exploitation counts come from operator records that were not independently verified.
What Was Handed to AI Agents, and What Was Not
This is the distinction that matters for the headline. STRU assesses with high confidence an “AI-executed, human-approved operating structure.” A human operator started each campaign with a victim-specific master prompt and reviewed every step under the directive “coordinate every action with me.” The AI stack, OpenCode as the interface, commercial models hosted on OpenRouter (STRU names anthropic/claude-opus-4.6 and moonshotai/kimi-k3 among those used), Phantom’s skill library, and a persistent knowledge base, planned reconnaissance, selected techniques, generated commands, triaged stolen data, and drafted reports and extortion material. “WormGPT” was a persona prompt and a set of MCP servers, not a custom model. When a model refused a request, STRU says the operator rewrote it as an “Authorized Red Team Assessment Request” instead of switching models.
Atos TRC is even more explicit. Its conclusion states that “the recovered evidence does not support the idea of autonomous AI conducting intrusions on its own.” AI “acted as a force multiplier for human operators,” reducing the effort to research targets, develop scripts, analyze data, document findings, and coordinate campaigns. Atos also notes that Phantom coordinated workflows, while individual actions could still run through custom Phantom agents, Mythic or Apollo C2, tunneling tools, or manually operated utilities. Recovered operator prompts Atos quotes are short and human-written, for example asking the assistant to continue from a session backup or to escalate from a domain user to Domain Admin given a service list. So the sheet headline “Handing Intrusions to AI Agents” overstates it. The accurate framing is that affiliates handed parts of the intrusion, planning, command generation, data triage, and drafting, to LLM-driven agents, while a human still chose targets and approved steps.
Separately, both teams describe plenty of plain automation and conventional tooling that is not agent decision-making: FOFA-style scanning of edge devices, reverse-SOCKS tunnels (RevSocks, Chisel, Ligolo), scheduled-task persistence, BYOVD security-tool killers, restic or rclone exfiltration to cloud storage, and custom INC encryptor builds. STRU reports operator records claiming large-scale abuse of FortiClient EMS (CVE-2026-35616), SimpleHelp (CVE-2024-57727), and cPanel (CVE-2026-41940), and labels those counts as operator-reported. Treat the scanning and exploit claims as operator data, not confirmed victim totals. STRU also recovered 46 personalized INC encryptor builds staged for an October 1, 2026 wave and a September 30 SimpleHelp screenshot showing encryption on a workstation; deployment against the staged targets is unconfirmed.
What Changes for Defenders
The recommendations below are my opinion, not vendor guidance and not legal advice. If the planning and typing are machine-assisted, your response timelines need to assume a shorter dwell between foothold and impact.
- Re-measure the first-alert-to-impact window. Build tabletop scenarios that assume continuous command generation overnight, not a human who sleeps. The SOCRadar and Atos material both show operators active on a schedule, but the agent does not get tired between approvals.
- Prioritize edge and RMM hygiene. STRU’s hunting notes focus on rogue EMS admins, modified VPN on_connect scripts, unauthorized SimpleHelp or MeshCentral agents, and MSP credential reuse. Patch those surfaces and require MFA and IP allowlists on every remote-management path.
- Alert on tooling the agents call. Watch for RevSocks, Chisel or Ligolo patterns, restic or rclone to unexpected cloud storage, Zemana-style vulnerable drivers paired with security-product termination, and mass .INC renames or INC-README.txt drops.
- Treat data-extortion clocks separately. These operators preferred proof packs, pricing from stolen financials, and customer-facing notice pages. Have IR, legal, and notification templates ready before encryption starts.
Picture this hypothetical. Your overnight SIEM shows a new EMS admin and a burst of PowerShell that looks like a junior analyst scripting. Under an agent-assisted playbook, that burst may be the middle of an approved kill-chain step, not a tired human. That scenario is my illustration, not a case from either report.
The Bottom Line
Exposed affiliate infrastructure shows LLM agents planning steps, generating commands, and drafting extortion text inside a human approval loop. It does not show fully autonomous ransomware. In my view, the practical change is shorter assumed dwell and louder hunting around the tools those agents call, not a rewrite of every detection rule overnight.
