Ransomware affiliates used AI agents to target 62 organizations
A person approves each step, but AI models from other companies do the intrusions. The operator rewrites an instruction that a model refuses as a security test.
Researchers at SOCRadar found an exposed server. It shows how affiliates of the INC and The Gentlemen ransomware groups use AI. A framework with the name Phantom does the intrusions with AI models from other companies. A human operator gives an instruction for each victim and examines each step. The logs show 8,257 messages in 133 sessions. SOCRadar estimates that 62 organizations were targets.
How it works
Phantom runs through OpenCode, a coding tool, with models that OpenRouter hosts. If a model refuses an instruction, the operator rewrites it as an approved red team test. The operator does not change to a different model. AI knowledge bases monitor victims, stolen credentials and the condition of the campaign.
The numbers
- SOCRadar estimates 62 targets. 4 are on leak sites.
- The US has more than 50% of the targets.
- The deadline to pay is 24 to 72 hours.
One password that persons use again opened the VPN, servers and security tools.
What is not known
The result of the ransom demands is not known. Researchers found no cryptocurrency wallet and no payment record. The logs show mixed results against enterprise endpoint security software. SOCRadar says the actor speaks Russian.
Insight: Attackers rent the model and bypass its limits with words
The operator rewrites an instruction that a model refuses as a red team test. Symantec says on 21 October that a different agent uses stolen access to the AI services of other persons. The victim pays. Attackers seem to rent the model and its bill from other companies, and they bypass its safety limits with words. The model does not stop them. The systems they go into do.
Sources
Posted
