A Python script named deadman.py, sitting in an open directory used by the Qilin ransomware crew, reads like something built to a specification. It waits for an appointed time, pushes a wiper to every Windows machine in a domain through Active Directory Group Policy, and narrates itself as it goes: stage the wipe payload to SYSVOL, stage the startup script, create the Group Policy Object on the domain controller. Cisco Talos says, with medium-to-high confidence, that a large language model wrote it.
Criminals using AI is not new. Reading the leftovers in their source code is a different kind of evidence.
The comments they forgot to delete
Talos found the directory while investigating an environment hit by Qilin, and three scripts carried the same fingerprints. veeam_kill.py stops, disables and destroys Veeam backups, its main function split into four numbered steps with progress logs at a uniform level of detail. deploy_locker.py, which pushes the encryptor out to multiple endpoints, opens with documentation style text laying out the tool's purpose, prerequisites and usage examples, the shape code takes when a model is handed a spec. The shell history on the same machine records somebody inspecting a directory belonging to a tool called llm_chatbot.
None of that proves a model wrote the encryptor. It does suggest the operators lean on one for the plumbing, and the plumbing is where an affiliate's hours go. It joins a growing set of cases in which criminals rent an AI built to write malware on request.
Inside The Gentlemen's attack server
Talos also identified open directory infrastructure it believes belonged to an actor tied to The Gentlemen. The contents amount to a complete kit: RustHound for mapping Active Directory, Responder and Impacket for relaying Windows logins, Ligolo-ng and chisel for tunnelling into victim networks, AnyDesk for hands on control, Rclone for moving data out, and the open source AdaptixC2 framework running command and control.
The bash history reads like a diary. The operator swept exposed hosts with masscan and nmap, exploited CVE-2025-24799, an unauthenticated SQL injection flaw in GLPI, and left traces suggesting attempts at Zerologon (CVE-2020-1472) and MS17-010. To take the data they mounted a backup share, opened the VHDX images inside, copied out ntds.dit, SAM and SYSTEM, ran secretsdump.py against them, and pushed the results to Wasabi cloud storage through rclone.
Then there is the keyboard. A keepalive script on the server carried Russian language comments, and the history holds strings such as "црщфьш" and "ды", which are "whoami" and "ls" typed with a Russian layout active. Talos reads that as an indication a Russian speaker was involved, consistent with existing suspicion that the group is led from Russia. The Gentlemen has run a ransomware as a service operation since roughly July 2025, claimed Air Canada data last week, and lifted its worldwide leak site listings from 48 in January to 105 in July.
Japan's attackers are moving down market
The headline count for Japan barely moved: 90 organizations affected from January to July 2026 against 86 a year earlier, up about 4.7%. What changed is who gets hit. Organizations capitalized at under JPY 1 billion accounted for 78% of victims, up from 69% in 2025, and manufacturing absorbed 34% of incidents. The Gentlemen led the table with 14, Qilin and SafePay with seven each.
Close the doors they actually used
The intrusions started at exposed VPNs, remote access services and network devices, or with stolen credentials. Inventory what is reachable from the internet, patch it or retire it, restrict management interfaces by source address, and put multi factor authentication on VPNs, cloud services and every administrative account, including those of subsidiaries and third party vendors. Snort coverage ships as SIDs 1:67111 and 7:29. The report is by Takahiro Takeda, Jordyn Dunk and Michael Szeliga, published by Cisco Talos.
This briefing is provided by IntelFusions for informational and defensive purposes only. It is based on sources assessed to be reliable at the time of writing, and analytic judgments carry the confidence levels indicated. Indicators of compromise are defanged; re-arm them only in controlled environments. IntelFusions is not affiliated with the organizations named and makes no warranty as to completeness or accuracy.