The cybersecurity industry is currently captivated by the emergence of autonomous, agentic offensive frameworks. With recent reporting from Google, CrowdStrike, Microsoft, and the FBI have all been highlighting the operational use of tools like ARTEX, Cairn, CyberStrikeAI, and PentAGI, among others. The prevailing narrative assumes every cyber threat actor is racing to automate intrusions from end to end.
From my vantage point of tracking ransomware campaigns daily, however, a striking contradiction emerges. While lower-tier crews and emerging actors are experimenting with agentic tooling, top-tier Ransomware-as-a-Service (RaaS) affiliates, like Akira, Qilin, or Rhysida, are conspicuously absent from this trend.
Understanding why requires looking past the hype of autonomous exploitation and examining the actual operational, economic, and architectural realities of running ransomware attacks. This article was also inspired by me recently pondering to the Twitter/X hivemind about why we are not seeing all of the top ransomware gangs using ARTEX and other AI pentesting tools.
Experimenters vs. Established Affiliates
Newer outfits and smaller extortion groups are visibly testing the waters. Researchers at CtrlAltIntel have seen WhiteArmor, for instance, leveraging ARTEX, though with notoriously poor operational security, operating the framework from the exact same VPS infrastructure tied to their original clearnet domain. Similarly, I have observed the likes of Settra utilising generic AI tooling. Thomas Roccia has also observed a newer gang called n0n using AI coding agents running obliterated models to assist their operations. Researchers from SOCRadar noted that n0n spent over $7k USD in paid LLM API usage to do so.
Despite the hype surrounding agentic frameworks, there are very little public reports from the well-known DFIR firms mentioning the adoption of ARTEX, Cairn, or PentAGI among elite ransomware affiliates. It seems, for established operators running multi-million-dollar intrusion pipelines, autonomous AI frameworks do not solve the right problems and introduce unacceptable operational risks.
Why Top Affiliates Are Passing on Agentic Frameworks
High-tier ransomware affiliates target major enterprises protected by mature, 24/7 Security Operations Centers (SOCs) and aggressive Endpoint Detection and Response (EDR) tooling. In these environments, stealth and precision dictate whether an intrusion results in a seven-figure pay-out or immediate containment. When an affiliate already possesses battle-tested, manual tradecraft and custom tooling that reliably bypasses modern defenses, delegating control to an erratic AI agent is a liability, not an upgrade.
The current generation of offensive AI frameworks (especially ARTEX) is heavily skewed toward external reconnaissance, web application exploitation, and initial access. However, top ransomware affiliates rarely bother with initial access themselves. Top affiliates routinely purchase pre-vetted corporate access directly from Initial Access Brokers (IABs) on underground forums.
Once inside, their core competency is rapid internal post-exploitation: Active Directory enumeration, lateral movement, privilege escalation, data exfiltration, and hypervisor encryption targeting Windows domains and VMware ESXi environments. Agentic frameworks currently offer little value in high-speed Windows/ESXi post-exploitation, leaving no compelling use case for the operators who execute the final stages of an attack.
Tokenomics, Hardware Sanctions, and OPSEC
For ransomware gangs, running autonomous agents at scale introduces severe infrastructure and financial headaches. Routing intrusion commands through commercial cloud LLM APIs risks logging, account suspension, and attribution exposure to defenders and law enforcement. Lesser groups with lax operational security (OPSEC) may tolerate this risk; top affiliates won't.
To maintain OPSEC by avoiding cloud exposure requires self-hosting capable models locally. Beyond the engineering overhead of maintaining dedicated GPU clusters and retraining crews on new workflows, hardware acquisition presents a major bottleneck. For Russian-speaking cybercrime syndicates operating under strict export controls, importing high-end enterprise GPUs is prohibitively expensive and logistically fraught, such as Ukrainian drone strikes on Russian data centers. Combined with the massive token consumption required for iterative agentic loops, the return on investment simply does not beat traditional manual operations.
What They Are Actually Doing
The absence of ARTEX in major ransomware incident response reports does not mean top affiliates are ignoring AI entirely. Rather, they are likely utilising locally hosted and uncensored models behind the scenes to augment human operators. Using local LLMs for tool and script writing, code obfuscation, phishing lure generation, or rapid parsing of exfiltrated documents leaves little-to-no distinct signatures.
Unlike an autonomous agent hammering a perimeter or stumbling through an internal network, local operational assistance is virtually invisible to us without some sort of OPSEC mistake exposing them.
The Only Known Cases
Security researchers from SOCRadar have observed one affiliate of INC/Lynx ransomware using CyberStrikeAI as an orchestration layer to attack FortiGate devices using unpatched CVEs and once inside, reportedly it was used to target Active Directory and VMware ESXi systems. However, the researchers noted it cost them $4.5k USD on OpenRouter to do so.
In another case, SOC Radar uncovered an affiliate of INC and TheGentlemen running intrusions through Phantom, a new Autonomous Red Team Kill Chain Framework. It reportedly uses OpenCode and Claude Opus/Kimi, wired to 32 MCP servers covering everything from recon to extortion, including a "ransom-ops" server with negotiation and pricing tools. Usage of the tool reportedly cost them $23.9k USD on OpenRouter to do so.
INC/Lynx and TheGentlemen are long-running established RaaS operations and have victimised many victims, over 1,000 organisations have been listed on each leak site. Therefore it is interesting to note that there is at least a couple of cases where top affiliates are using AI but it has not been widely observed so far, considering how long these autonomous pentesting tools and obliterated models have been available.
Looking Ahead
For now, the reluctance of top ransomware affiliates to massively adopt agentic frameworks seems to be a rational business decision. Their existing tools and TTPs remain overwhelmingly effective, their division of labour with IABs insulates them from the problems AI currently solves, and their aims of hitting hard targets forbid the noisy, unpredictable behavior of early autonomous tools.
Until agentic frameworks master quiet, reliable Windows and ESXi post-exploitation, and become cheap enough to run covertly, the elite tier of ransomware will likely continue to let smaller, sloppier crews beta-test the future of automated attacks.
Additional Resources
- Sky News' Rowland made a great video breaking down why there isn't really an AI cybergeddon. My advice is to send this to any executives with AI psychosis.
- SANS uploaded a great podcast with Rebecca Taylor who infiltrated the cybercrime underground and shared what she saw from cybercriminals talking about AI.
Introduction to Malware Binary Triage (IMBT) Course
Looking to level up your skills? Get 10% off using coupon code: MWNEWS10 for any flavor.
Enroll Now and Save 10%: Coupon Code MWNEWS10
Note: Affiliate link – your enrollment helps support this platform at no extra cost to you.
Article Link: Top Ransomware Gangs & AI Pentesting Tools
