theaisafemode.com
AI Safe Mode

Ransomware Crew Let AI Agents Run Its Break-Ins

An old rusty padlock and bolt on a wooden door; illustrative
Illustrative photograph only, not an event image. Photo by Cristina Gottardi via Wikimedia Commons; CC0 1.0. See LICENSES.md.

Exposed attack servers showed a ransomware crew used AI agents to run break-ins, with a human approving each step: 8,257 AI messages, 62 organisations named.

Researchers found exposed attack infrastructure used by ransomware affiliates that held thousands of AI chat messages showing how the crew used AI agents to plan and run break-ins, with a human approving every step.

According to The Hacker News (Oct 23, 2026) and Hackread (Oct 21, 2026).[1] [2]

What happened

A threat research team found exposed attack infrastructure linked to affiliates of two ransomware operations.[1] [2] Hackread reports it was seven exposed servers, holding AI session logs, stolen credentials and ransomware tools.[2] Hackread reports the findings were published in a report on Oct 20, 2026.[2] Hackread covered them on Oct 21 and The Hacker News on Oct 23.[1] [2]

According to The Hacker News, reused passwords did much of the damage: in one case, a single reused password unlocked the VPN, servers and backups at once.[1]

The researchers recovered 8,257 AI messages across 133 sessions in an AI coding tool.[1] [2] They also found 46 custom encryptor builds prepared for six victims.[2]

The records showed more than $23,920 spent on commercial AI services.[2]

How the crew used AI

The operators used an AI coding interface, plus AI-managed knowledge bases that tracked victims, stolen credentials and campaign progress, to plan reconnaissance, write commands and draft reports.[1] [2]

The AI did not act alone. The researchers describe an “AI-executed, human-approved” set-up: a person gave a victim-specific prompt and reviewed each step.[1] One instruction told the AI to “coordinate every action with me.”[2]

When the AI models refused harmful requests, the operators reworded their prompts as if they were an authorised security test, pretending to be legitimate testers, rather than switching to a different model.[1] [2]

The exposed data also included credentials stolen during the intrusions, which the operators fed to the AI to analyse and decide where to go next.[2]

The servers also held a custom framework for coordinating attacks, command-and-control tools, and software designed to switch off security products. Hackread also lists breach and target counts the attackers claimed for three known software flaws, and notes that those figures come from the attackers’ own records and have not been independently verified.[2]

Who was targeted

The crew aimed at internet-facing systems and remote-access setups, the doors companies leave open so staff can work from anywhere.[1] [2]

The operators’ records named 62 organizations, more than half of them in the United States.[1] [2] Hackread puts the US count at 36.[2] According to Hackread, industrial and manufacturing firms made up 16.1% of the list, followed by healthcare and pharmaceutical firms at 14.5%.[2]

The encryptor builds were prepared for a planned Oct 1, 2026 deployment. The report doesn’t confirm that all six attacks happened, but a Sept 30 screenshot showed an encryptor running on a victim’s computer in at least one case.[2]

The plan for extortion was built around stolen financial and business data, with threats to contact victims’ customers, partners and regulators.[1] [2] The Hacker News notes the final outcome of the ransom demands is unknown, because no payment record was recovered.[1]

Why it matters

This is a rare look inside a working criminal crew using AI agents during real intrusions. The AI speeds up the dull parts of an attack, such as scanning, scripting and sorting stolen logins, while a human keeps control.[1] [2]

It also shows that the “safety refusals” built into AI models can be talked around by dressing a request up as an authorised security test. Defenders can’t count on the AI saying no.[1] [2]

The short version

The crew did not need a secret super-tool. They paid for ordinary AI services, dressed their requests up as an authorised security test, and let the AI do the boring work of an attack while a person approved each move. The weak points they went after were the usual ones: open remote-access systems and stolen passwords. Closing those doors still stops most attacks, whether or not AI is helping the attacker on the other side.

How to stay safe

  • Never reuse passwords at work. This matters most for VPN, remote-access and backup systems.
  • Turn on multi-factor sign-in. Use it for every remote-access and admin account.
  • Patch internet-facing systems fast. Switch off remote-access tools nobody uses.
  • Keep an offline backup. Make sure ransomware can’t reach it, and test restoring it.

Follow AI Safe Mode

Protect yourself with Privaku

See the post on YouTube Watch and follow AI Safe Mode on YouTube @AISafeMode and Instagram @aisafemode.

References

  1. The Hacker News, Oct 23, 2026. Report on the exposed AI-driven intrusion setup. Read the report ↗
  2. Hackread, Oct 21, 2026. Report on the exposed servers and AI chat logs. Read the report ↗
Facts come only from the published reports above and are attributed to them. AI Safe Mode doesn't name companies, products or people involved in incidents. Spotted an error? Email [email protected] and it will be reviewed and corrected. See the corrections policy.