What happened in the Aurora campaign?
Two research teams analyzed the same mistake: an Aurora affiliate reportedly left a server exposed to the open internet.
Gambit Security, a Tel Aviv threat-intel firm, says it recovered 28 chat sessions between Aurora operators and Cursor's AI agent, covering hands-on exploitation across ten target organizations from April 8 to May 21. Per Gambit, the operators falsely framed the work as a simulation, and the agent went on to run hundreds of malicious operations, including credential theft and account takeover. Reuters reviewed portions of the chat data and identified six victims.
CloudSEK examined the same server and counted more than 20 targeted organizations across nine countries between April and July.
Both teams agree on the point most coverage will blur: a human ran these intrusions. CloudSEK states plainly that nothing in the evidence shows the agent acting on its own. The operator used AI to plan and troubleshoot, and made every call.
Did AI change how ransomware works?
Not the tactics. The tradecraft in the logs looks like any competent intrusion from the last five years. What changed was pace.
Sela, Gambit's director of threat intelligence, estimates the agent made the crew 30 to 50 percent faster by handling work they would otherwise have done by hand. CloudSEK reads it the same way: general-purpose AI helped the operator plan recon and sequence known attack paths more quickly.
Every defense assumes some amount of time to notice, respond, and recover. The Aurora logs say attackers now spend less of it.
Why does the payload go after recovery first?
CloudSEK's teardown describes a Windows payload that enabled backup privileges, deleted Volume Shadow Copies, resized shadow storage to purge older copies, and disabled System Restore, all before encrypting anything. The victim list reportedly included IT and backup infrastructure firms, so the recovery layer itself may have been among the targets.
A victim who can restore cleanly has no reason to pay, so attackers kill restore paths first. The cost of losing that race shows up in Sophos's 2026 State of Ransomware report: average recovery ran about three weeks and $1.7 million.
Cloud teams see it coming. In Eon's 2026 survey of 583 cloud infra leaders, 77% said they worry their recovery environments could be targeted, and 63% said they discover protection gaps only after an incident.
What does recovery that survives a faster attacker look like?
- Recovery copies in a separate trust domain. Keep them in a different account or organization, with credentials production never touches. A second region is replication, not isolation: if the same credentials reach both, both fall together.
- Immutability the platform enforces. A retention lock an admin can flip is a lock an attacker with admin credentials can flip. Our Gunra teardown covers how two data centers trusting the same credentials lost their copies together.
- A clean restore point, at record level. Ransomware crews typically sit in an environment for five to seven days before deploying, so restoring last night's copy can restore the intruder with it. You need to see where tampering started and roll back only those clean records.
- A measured restore time. Most teams first test a full restore mid-crisis, which is where the three-week averages come from. Run drills often so you’ll have the right strategy in place in the event of a real scenario.
Where Eon fits
Eon Data Protection keeps recovery copies in an immutable, logically air-gapped vault that no production credential can reach. Setup runs through one read-only IAM role, with no appliances to install. The platform detects ransomware and rogue-agent tampering across your whole cloud estate, including inside your databases. It rolls back exactly what the attack affected, down to individual records, in minutes.
The speed holds up on real workloads: NETGEAR cut recovery on a 10TB SQL Server workload from 24 hours to 3 hours, 88% faster.
If the Aurora story prompts a posture review, our cloud ransomware guide walks through the full checklist.
FAQ
Did the AI agent carry out the Aurora attacks on its own?
No. A human operator ran the intrusions and used the agent to plan and troubleshoot. CloudSEK states the evidence shows no independent action by the agent, and Gambit describes the operators steering it session by session.
What is Volume Shadow Copy deletion?
Volume Shadow Copy deletion: removing Windows' built-in point-in-time snapshots (VSS) so a victim can't roll files back locally. Ransomware has led with the move for over a decade because it pushes victims toward slower external restores, or toward paying.
How much faster does AI make ransomware crews?
Gambit's Eyal Sela estimates 30 to 50 percent for this affiliate, based on the recovered sessions. Treat it as a single measured campaign rather than an industry constant; Reuters noted that it could not independently determine how much the agent contributed to each breach.
Do immutable snapshots stop attacks like this?
Only when the lock sits beyond the attacker's reach. Retention settings an admin can change offer little once admin credentials leak, so the copies belong in a separate account with platform-enforced locks. Our AWS immutable backups guide covers the setup.
How fast do most organizations recover from ransomware?
Slowly. Sophos's 2026 report found only 16% of victims recovered in under a day. In Eon's 2026 survey, 60% of cloud IT leaders said a full restore takes six hours or more.
