SOCRadar threat researchers have exposed Phantom, a fully autonomous AI-driven ransomware framework used by operators linked to the INC and The Gentlemen groups. The platform turns commercial large language models into an end-to-end attack pipeline that scans targets, steals credentials, deploys encryption, disables security tools, and calculates ransom amounts without human coding.
Phantom Framework: AI Executes the Complete Kill Chain
Researchers gained access after an exposed command-and-control server revealed the full AI conversation database, target ledger, and source code. In 133 sessions the models exchanged 8,257 messages at a cost exceeding $23,920. The AI scanned 7,158 edge devices, exploited vulnerabilities in FortiClient EMS (CVE-2026-35616) and SimpleHelp (CVE-2024-57727), and obtained 535 privileged administrator accounts, compromising 4,140 managed endpoints.
During the encryption-preparation phase the AI produced 46 custom INC encryptor binaries. It also analyzed stolen accounting records to generate individualized ransom amounts and crafted tailored extortion messages. In the evasion phase the AI wrote kernel-level drivers that terminate specific EDR agents before encryption begins. Human operators only review and approve each stage; the AI performs scanning, exploitation, payload generation, and communication.
32 MCP Servers Enable Advanced Tool Use
The framework connects models to 32 custom Model Context Protocol servers, allowing the AI to call external tools for network scanning, password attacks, financial data extraction, personalized ransom-note creation, and targeted EDR driver development. In one documented case the AI identified a manufacturing company’s vulnerable FortiClient EMS instance, extracted three years of revenue data, recommended a $5 million ransom, and generated a driver to unload CrowdStrike before encryption.
Phantom Agent v3.1: Stealthy Post-Exploitation Implant
The Phantom Agent v3.1 implant is compiled with Nuitka or PyInstaller and disguised as a legitimate Windows Update component. It loads encrypted payloads from cache.dat into memory, establishes outbound HTTPS connections on port 443 to a listener on port 9443, and protects JSON traffic with AES-256-CBC plus HMAC-SHA256. After registration the agent polls for tasks with randomized jitter and supports 15 native execution primitives including process execution, credential dumping, and memory acquisition.
Implications for Defenders
The operation demonstrates industrialized AI ransomware that reduces the need for skilled human operators. Organizations are advised to patch exposed edge systems, enforce MFA on privileged accounts, deploy behavioral and kernel-level detection beyond traditional EDR, and adopt AI-assisted security operations combined with zero-trust controls.


