Aur0ra’s operators have been outsourcing the legwork to AI agents – and it worked well enough to reach domain admin on some networks.
Sophos, which tracks the crew as GOLD OGRE, recovered the evidence from Cursor, the coding assistant, dating to late August. Operators fed the agents VPN credentials, proxy access, target lists and objectives, then largely left them to it. The agents mapped networks and Active Directory, trawled shares, documents and backups for passwords, and pivoted laterally. Three breached organisations ended up named on Aur0ra’s leak site; a handful of networks saw the agents take domain-admin rights.
Entry and encryption remained human work. The typed instructions that survive are revealing. One operator flagged a Windows AD environment for testing and offered the agent user accounts plus a SOCKS5 relay. Another called off brute-forcing domain administrator accounts in favour of quieter routes. A third wanted backup servers, NAS systems and virtualisation hosts located. Cyrillic keyboard slips in the logs suggest Russian-speaking operators, the company says.
To slip past built-in guardrails, the crew framed the work as authorised red team activity or diagnostic testing. Cursor’s Composer agentic model did most of the running, with systems from Anthropic, xAI and Moonshot alongside; Anthropic’s share fell as toolmaking gave way to intrusive operations, which Sophos ties to shifting objectives.
Sophos’s Counter Threat Unit is highly confident the group tested with AI whether its tooling could evade EDR, then shared the technique with affiliates. The researchers expect other crews to adopt agents too. Two further outfits, Gambit Security and CloudSEK, have also written up Aur0ra, and both appear to have worked from the identical exposed open directory.
