Aur0ra attackers used Cursor agents to reach domain administrator access

Sophos said it identified the activity while investigating an AI-assisted testing framework built with Cursor.[1] Its Counter Threat Unit published research on the framework in May 2026, and Sophos said the framework had tested multiple endpoint detection and response products, including its own agent.[1] Researchers later found the same infrastructure hosting Cobalt Strike for Aur0ra operations.[1] Sophos researchers said with high confidence that GOLD OGRE used AI to test tools’ ability to evade detection and then supplied similar technology to Aur0ra associates.[1] The reporting links tool testing and criminal operations, but does not establish that every test was part of an intrusion.

CyberSecurityNews reported that human operators supplied VPN access, targets and goals, while Cursor AI agents handled most of the intrusion work with little assistance.[1] The agents mapped networks and checked for weak paths to privileged accounts.[1] Attackers tried password spraying, Kerberoasting, NTLM coercion and relay, and credential extraction from files and backups.[1] Sophos said agents reached domain-management access in some victim networks; three of those victims later appeared on Aur0ra’s leak site.[1] The report also says attackers described the work as authorized red-team work or diagnostic testing to get past AI safety rules.[1]

CyberSecurityNews reported that Cursor’s Composer model handled many intrusions alongside models from Anthropic, xAI and Moonshot.[1] It also reported that the group initially relied almost entirely on Anthropic and added other models as it shifted from tool development to network intrusions.[1] Separately, Gambit Security reported observing Cursor Agent running Claude Sonnet in hands-on exploitation against ten target organizations between April 8 and May 21, 2026.[2] Gambit found that most agent commands did not achieve their stated objective on the first attempt, prompting repeated refinements to commands and scripts.[2] Its investigation also uncovered a Linux ransomware variant capable of targeting ESXi environments.[2]

Sophos said it warned affected customers and worked with law enforcement to notify other victims.[1] Its recommendations include patching internet-facing devices, using phishing-resistant multifactor authentication and closely monitoring endpoint detection and response alerts.[1] For limiting intruder movement and protecting recovery options, Sophos also recommends segmenting critical servers, isolating backups, reviewing domain administrator rights, using privileged access management and applying application controls.[1] These measures address the access and data-handling risks described in the reporting; the evidence does not say how many organizations have adopted them.