malware A field note by Vikrant Sharma
Malware that scores victims with LLMs
The Volta infostealer uses a language model to rank stolen data by victim value before deciding what to exfiltrate. No human triage needed.
The Volta infostealer does something I had not seen before. It runs a language model on the infected machine to score the victim before sending data back to the attacker. Most infostealers grab everything and sort it later. Credentials, browser cookies, crypto wallet files, all of it gets exfiltrated to a command-and-control server. The attacker then manually filters for high-value targets. That takes time and bandwidth. Volta flips this. After harvesting credentials and files, it feeds metadata about the victim into a local LLM. Job title from email signatures, browser history showing SaaS admin panels, presence of corporate VPN configs. The model outputs a score. If the score is high enough, the malware exfiltrates the full dataset. If not, it discards most of it and moves on. This is efficient. The attacker saves bandwidth by not hauling gigabytes of low-value data across the network. They also reduce their exposure window because high-value victims get exploited faster, before incident response teams catch the infection. The model runs entirely offline. No API calls to OpenAI or Anthropic. Volta uses a quantised model small enough to bundle with the malware binary. Talos Intel reverse-engineered the scoring logic and found it weighted things like admin credentials, financial keywords in documents, and evidence of privileged network access. What surprises me is how little compute this needs. A 7B parameter model quantised to 4-bit runs on a laptop CPU in seconds. The malware does not need a GPU. It scores a victim faster than most EDR tools detect the initial infection. This is not theoretical. Talos observed it in the wild targeting corporate employees. The attacks are selective, which makes them harder to spot in aggregate. No mass spam campaigns, just targeted phishing with a smarter second stage. Defenders now have to assume the malware is smarter than a dumb file-grabber. If your detection relies on seeing large outbound transfers of credential files, this bypasses that. The exfiltration only happens after the malware decides you are worth it.
Source: Malware uses LLMs to rank victims by how much they’re worth
