Ransomware gang GOLD OGRE used AI agents in victim networks

Sophos says the group behind the Aur0ra ransomware used AI agents in victim networks. The agents collected passwords and got domain administrator access. Sophos tracks the group as GOLD OGRE. It found files from the Cursor coding assistant from late August. The files show the attackers giving the agents VPN logins, proxy access and targets.

How it works

The agents mapped each network. They searched shares, documents and backups for passwords. They moved to other systems with small input from the operators. People kept control of the break in and of the ransomware deployment. To get around guardrails, the attackers said the work was approved red team testing.

The proof

Sophos names the Composer model of Cursor and models from Anthropic, xAI and Moonshot. It says with high confidence that the group used AI to test if its tools could evade endpoint security products. Keyboard slips in Cyrillic suggest to Sophos that the operators speak Russian.

What is not known

Gambit Security and CloudSEK published their own analyses, apparently from the same exposed open directory. Sophos thinks other ransomware groups will very probably add AI agents. It says patching, phishing resistant multi factor login and alert monitoring still work, but a quick response decides the result.

Insight: The red team cover story works on many models

Sophos says GOLD OGRE called its intrusions approved red team work. On 23 October, SOCRadar reported that other affiliates rewrite a refused instruction as an approved security test. On 31 October, Google reported one group that went from a first account to deep access in about four hours. This reads as a shared trick that a model cannot check, so refusals depend on a claim, and defenders win only on speed. Time your own response.