---
title: AI in the wild: ransomware operators got commercial models to run intrusions by relabeling requests as authorized red-team work | ctxsec
url: https://ctxsec.com/items/ai-socradar-ransomware-ai-co-operator-refusal-rewording
description: ""
status: "confirmed research (SOCRadar, Oct 23 update); analysis of attacker logs"
published: 2026-10-23
---

> 

[← AI](/ai) · [markdown](/items/ai-socradar-ransomware-ai-co-operator-refusal-rewording.md)

confirmed research (SOCRadar, Oct 23 update); analysis of attacker logs

2026-10-23

### AI in the wild: ransomware operators got commercial models to run intrusions by relabeling requests as authorized red-team work

The most useful AI finding in SOCRadar's INC and Gentlemen report is how ordinary the stack was: OpenCode as the agent shell, standard commercial models through OpenRouter (the logs name Claude Opus 4.6 and Kimi K3), a persistent knowledge base, and a library of 18 Phantom "skills". The AI planned recon, wrote commands, triaged stolen files and drafted extortion letters; a human approved every action under the directive "coordinate every action with me".

- Refusals were handled by rewriting prompts as an "Authorized Red Team Assessment Request", not by switching to an uncensored model, which shows why authorization claims inside a prompt shouldn't be trusted.
- The AI's memory became the evidence trail: session logs and knowledge bases recorded victims, credentials and progress, which is how SOCRadar reconstructed the campaign.
- A partner guide advertised handing the whole kit (Phantom, OpenCode setup, knowledge base, MCP servers) to other operators.

Sources: [SOCRadar STRU: INC and The Gentlemen Ransomware: Exposed AI Infrastructure (Oct 20, updated Oct 23)](https://socradar.io/blog/inc-gentlemen-ransomware-ai-infrastructure/)