Last week the agentic intrusion story was nation-state crews and ARTEX. This week it's a ransomware gang. Sophos published research on GOLD OGRE, the group behind Aur0ra ransomware, and the short version is that the boring middle of a ransomware attack (discovery, credential hunting, lateral movement) is now being handed to AI agents.
What Sophos Found
- The agents did the grind. Human operators gave the agents VPN credentials, SOCKS proxy access, targets and objectives. From there, per Sophos, agents ran network and Active Directory discovery, analyzed credential attack paths, and moved laterally "with limited human involvement." The task list reads like a pentest checklist: password spraying, Kerberoasting, NTLM coercion and relay, credential extraction.
- It worked, sometimes. In a subset of environments, the agents reached domain management access. Sophos says three of those victims later showed up on Aur0ra's leak site.
- Humans kept the bookends. Initial access and the ransomware itself stayed under human control. The agents handled everything in between.
- The tooling was AI-built and AI-tested. Sophos says it watched an AI-assisted post-exploitation framework, developed with Cursor, being tested against several vendors' EDR products (Sophos's included), then saw the same infrastructure hosting Cobalt Strike for Aur0ra operations.
- Multiple models. Sophos says the intrusions ran through Cursor's Composer model plus models from Anthropic, xAI and Moonshot, starting almost entirely on Anthropic before adding the others.
- The guardrail bypass was a cover story. The operators framed the work as authorized red teaming or "diagnostics testing."
The source material came largely from an exposed open directory (Gambit Security and CloudSEK also analyzed it) holding Aur0ra binaries, shell history and Cursor chat logs. Attackers keep leaving their AI workbenches open on the internet. Cairn all over again.
Why This Matters
The scary part isn't that AI made the attack smarter. Kerberoasting isn't new. It's that the slowest, most skill-dependent phase of a ransomware intrusion just got cheaper and faster. Sophos says it directly: AI "can accelerate phases of the attack chain and reduce the level of technical aptitude needed to conduct an attack." Fewer skilled operators can run more intrusions at once.
That changes the defender's math in one specific way: your time between "they got in" and "they own the domain" is shrinking. If your detection and response plan assumes a human attacker poking around for days, that assumption is aging badly.
And the "it's for a red team" pretext is going to keep working on some models some of the time. Guardrails help. They are not your control.
What To Do
- Close the front door. Initial access was still human, and Sophos's first recommendations are the classics: patch internet-facing devices and deploy phishing-resistant MFA. VPN credentials were literally the first thing handed to the agents. The fewer things you expose to the internet, the fewer doors there are to hand over.
- Make lateral movement hard. Agents are great at enumerating what a stolen account can reach. Segment critical servers, and connect users to specific apps instead of putting them on a network where AD, backups and hypervisors are all reachable.
- Hunt the agent's checklist. Spraying, Kerberoasting, NTLM relay and mass share crawling are noisy. Alert on them, and make sure someone actually looks at EDR alerts quickly. Sophos is blunt that "successful defense relies on a quick response."
- Protect backups and virtualization. The operator prompts went straight for backup, NAS and virtualization servers. Isolate them and review who holds domain admin.
- Watch for AI tooling where it doesn't belong. An AI coding agent or a burst of LLM API traffic from a server that has no business running one is worth a look.
Bottom Line
Ransomware crews just automated the middle of the attack. You can't automate your way out with policy wording on someone else's model. What you can do is give the agent nothing to find: no exposed entry points, no flat network, no standing admin rights.
Sources: Sophos: AI used in Aur0ra ransomware testing and execution and Help Net Security