Ransomware gang used AI agents to move through victim networks, Sophos finds

What attackers presented to AI tools as authorized security testing was work inside ransomware victims’ networks, according to Sophos. In findings published October 29, the company described agents searching for credentials and moving between computers with limited human input. The operators delegated much of the work after gaining access—not the entire attack.

Credentials in, network access out

Sophos tracks the group behind Aur0ra ransomware as GOLD OGRE. Its analysts recovered Cursor coding-assistant artifacts in late August that exposed how operators handed tasks to AI. Humans supplied VPN credentials, proxy connections, targets and objectives. Agents then mapped networks and Active Directory, the system used to manage Windows accounts and access.

The agents did more than suggest commands. Sophos observed them executing commands, analyzing ways to exploit credentials and moving laterally—using access to one part of a network to reach others. They searched shared folders, user documents, backup repositories and cloud-connected storage for credentials. Their activity included attempts to obtain passwords and abuse network authentication.

In a subset of environments, the agents achieved domain-management access. Three of those victims were later named on Aur0ra’s data-leak site. Humans still controlled initial entry and ransomware deployment, as Help Net Security’s November 2 account of the findings emphasized. That boundary distinguishes the observed workflow from an attack run entirely by AI.

Don't try brute-forcing domain administrator accounts anymore. Try other, less noisy, methods and vectors.

Operator instruction recovered by Sophos, in the researchers’ English translation

A testing cover story, with human steering

The recovered instruction shows an operator redirecting an agent away from repeated password guessing toward less conspicuous approaches. Another prompt offered proxy access and user accounts to test a Windows network. A third requested a search for backup servers, network storage and virtualization systems—specific infrastructure targets rather than a general request to write software.

Sophos says the attackers bypassed AI guardrails by describing intrusions as authorized red-team exercises—security tests that imitate attackers—or diagnostic work. The operations used Cursor’s Composer agent and models from Anthropic, xAI and Moonshot. Anthropic models were used almost exclusively at first, before operators introduced others. Sophos suggests that shift may reflect a move from building tools to conducting intrusions; it does not establish the reason.

What the evidence connects—and what defense still requires

The investigation grew out of an AI-assisted framework for testing endpoint detection and response tools, which monitor computers for attacks. Sophos later observed the same infrastructure supporting Aur0ra operations. It assesses with high confidence that GOLD OGRE tested ways to evade detection, then provided associates with similar technology for ransomware attacks.

Earlier analyses by Gambit Security and CloudSEK apparently drew on the same exposed directory, Sophos notes. Those are not two independent collections of evidence. The directory contained ransomware files, command history and Cursor chat logs. Sophos found substantial overlap in techniques with its testing-framework investigation, but infrastructure overlap was limited to one proxy-related IP address.

Sophos says it notified potentially affected customers and worked with law enforcement to notify other organizations. Its defensive recommendations focus on blocking entry and restricting movement after a compromise, rather than treating AI involvement as a reason to abandon established controls:

• Patch internet-facing devices, use phishing-resistant multifactor authentication and respond quickly to endpoint-security alerts.
• Separate critical servers from other systems and isolate backups to limit attackers’ movement.
• Review domain-administrator privileges, manage privileged access and control which applications can run.