Cisco Talos said Saturday that a Windows infostealer calling itself VOLTA-AUTODOXER sends a structured bundle of stolen account data to large language models so operators can rank victims by estimated monetization value¹.

Researcher Ryan Fetterman wrote that the implant, surfaced through Talos's CAIRN effort to track AI-integrated malware, is self-deleting and does not rely on attacker-owned infrastructure¹.

Talos said VOLTA first harvests a typical stealer haul, then builds an evidence file that can include location, employer, card last-four digits, Slack workspaces and session-based enrichment from services such as Gmail and Discord¹.

That bundle is posted to OpenRouter under a system prompt that, Talos said, tells the model to act as a forensic intelligence analyst and to return ranked valuables and an assessment of the victim's security hygiene¹.

If a model refuses, VOLTA checks the reply for refusal language and can fail over across a chain of up to three models, which Talos said implies operators expect guardrail hits¹.

Inference runs on the victim's own machine and network before results are packed with the local identity report for exfiltration, Talos said; the LLM stage is optional if no API key is present¹.

Talos framed the design as effort displacement rather than a new theft technique: the model does triage that an operator would otherwise do by hand, and the same pattern may appear in other malware phases¹.

Independent confirmation of volume, attribution or active campaigns was not in the assignment. Talos's account is vendor research, not a government advisory¹.

Sources

1. Cisco TalosAppraisal by AI: Model-driven victim prioritization in the VOLTA infostealer10:29 AM

Read the original at Cisco Talos ↗