Incident sheet · /security

Filed
Read
4 min · 843 words
Severity
Routine

Ransomware Affiliates Spent $23,000 on AI APIs; Labs Stayed Silent

SOCRadar exposed an INC and The Gentlemen AI-driven intrusion ecosystem using Claude and Kimi K3 via OpenRouter — $23,000 in API spend, 62 targets, no lab detection.

AI labs and ransomware: Who’s responsible for catching misuse?
FeedAI labs and ransomware: Who’s responsible for catching misuse? — katielwhite91 / pixabay

Event log

  1. SOCRadar logged over 8,000 AI messages across 133 sessions targeting at least 62 named organizations.
  2. One attributed OpenRouter account spent more than $23,000 on commercial AI API access.
  3. Flashpoint estimates initial access costs on the ransomware black market fell 69%.
  4. The FBI's Internet Crime Report 2025 recorded 63 new ransomware variants last year.
  5. Acronis estimates 80% of RaaS operators promoted AI and automation features to affiliates.

A single attributed OpenRouter account tied to ransomware affiliates ran up more than $23,000 in commercial AI API spend, according to SOCRadar — and the abuse was exposed by a third-party security vendor, not by the AI labs whose models were used. The finding sits at the center of a growing dispute over whether frontier AI developers should detect criminal use of their own models or continue to rely on outside researchers.

SOCRadar's threat research unit detailed the campaign in a recent threat advisory. Affiliates of INC Ransom and the "The Gentlemen" ransomware-as-a-service (RaaS, a franchise model where operators lease malware to affiliates) operation built what researchers call an "AI-driven intrusion ecosystem" around publicly exposed command-and-control infrastructure.

What did SOCRadar actually find?

The attackers ran intrusion campaigns through OpenCode, an open source AI coding agent, using commercial models including Anthropic's Claude and Moonshot AI's Kimi K3. Access came through OpenRouter, an API key gateway, rather than directly from the labs.

The group also built its own tool, Phantom, described as an "autonomous red team kill chain framework." Phantom tracked victims and stole credentials. SOCRadar logged more than 8,000 AI messages across 133 sessions targeting at least 62 named organizations.

SOCRadar assessed the ecosystem with high confidence as "AI-executed, human-approved": the human operator directs and approves, while the AI stack plans reconnaissance, selects techniques, generates commands, triages stolen data and drafts reports.

The exposure was accidental in a broader sense. Had the operators' servers remained secured, the abuse might never have surfaced. It is not the first such case: in July, Cybernews researchers found an exposed server belonging to a Gentlemen affiliate holding 3.1 TB of stolen data and an agentic AI harness (an orchestration layer that lets models execute multi-step tasks autonomously) running automated attacks.

How is AI changing ransomware economics?

Large language models have shifted ransomware cost structures in measurable ways:

  • Flashpoint estimates the price of initial access to victim systems has fallen 69% on the ransomware black market.
  • Acronis researchers estimate 80% of RaaS operators actively promoted AI and automation features to affiliates and customers.
  • The FBI's Internet Crime Report 2025 recorded 63 new ransomware variants last year, with Akira, Qilin, INC Ransom/Lynx/Sinobi, BianLian and LockBit most frequently reported.

Dr. Hugh Thompson, program chairman of the RSAC Conference, described ransomware as a business with a sales funnel where AI is "collapsing the cost of the top of that funnel." The attack objectives — extortion, financial gain, reputational harm — have not changed. AI has changed the cost of reaching them.

"AI gives attackers scale, personalized lures for initial access and the ability to spin up infrastructure to execute attacks," Thompson told TechTarget. "The result is that a single operator can go after multiple targets at once and is also likely to have a higher success rate with those targets."

Steve Winterfeld, Advisory CISO at Akamai, said AI has produced "hyper-targeted phishing attacks on steroids" and that AI-driven automation is "industrializing the path to compromise and making it easier to scale criminal operations."

Why did the models comply?

SOCRadar's findings point to two obfuscation methods. Traffic routed through OpenRouter's gateway would not have appeared as ransomware activity to the labs. When models refused malicious requests, attackers reframed prompts as "Authorized Red Team Assessments," which appeared to result in AI compliance.

OpenRouter states that unauthorized "red teaming" can result in account termination. Anthropic and OpenAI both publish transparency reports on detecting, countering and handling misuse, and both maintain self-monitoring and refined guardrails. Anthropic has published reports on disrupting campaigns that exploit its models and on working with authorities, including findings from an AI-orchestrated cyberespionage campaign in November 2025. OpenAI also publishes disruption work but keeps details of law-enforcement cooperation vague.

Should frontier labs carry the responsibility?

There is little legal pressure to force the issue. California's SB 53 legislation requires frontier AI labs to report potential critical safety incidents, but a ransomware operator quietly accessing Claude through an API gateway would be unlikely to meet that reporting threshold.

The precedent is uncomfortable. Penetration testing software, built for ethical security testing, has long appeared in cybercriminal playbooks, and developers of those tools have rarely borne blame. Fortra cannot eliminate every cracked copy of Cobalt Strike in the wild, but it can seek court orders to take infrastructure offline. Frontier labs hold a comparable toolkit: account termination, transparency reporting, tightened guardrails and law-enforcement cooperation.

What should enterprises do in the interim?

Winterfeld argued the question of lab responsibility does not change enterprise exposure, and that minimizing impact when — not if — an attack occurs is among the strongest available measures.

"The best defense needs to be built around minimizing impact through AI-powered micro-segmentation to stop lateral movement once adversaries get in," Winterfeld said. "It stops the infection before it has a material impact. It's how you stop a perimeter breach from turning into a full-scale ransomware crisis."

via techtarget.com (Source)

Filed under

  • ransomware
  • ai-security
  • threat-intelligence
  • api-security
  • frontier-ai
Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Senior reporter covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.

199 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›