Windows malware VOLTA-AUTODOXER asks AI models to rank what it stole

What Talos found

As infosecu.technews.tw reported, Cisco's threat intelligence team Talos has analyzed a new Windows malware that calls itself VOLTA-AUTODOXER. It is an information stealer, meaning it harvests data from an infected computer and sends it to an attacker.

The loot list is familiar: browser passwords and cookies, credentials for logged-in services, credit card data, crypto wallets, and SSH and cloud-service keys. Saved Wi-Fi passwords are pulled out as plain text.

Building a profile of the victim

What sets VOLTA apart is the next step. It assembles a structured file on the person, including location and home address, the last four digits of a card, employer and company domain, an inferred birth year, and the author details embedded in documents.

It also uses stolen logged-in sessions to dig further, as if it were an investigator with the victim's accounts. It searches Gmail for receipts from Amazon, DoorDash, Uber Eats, Grubhub, PayPal and Walmart to confirm a home address, and it pulls order histories. It uses a stolen Discord token to fill gaps and checks Rolimons to estimate the value of Roblox items.

The AI step, and what it does

VOLTA then sends the profile to a model through OpenRouter, a service that routes requests to different AI models. The prompt casts the model as a "forensic intelligence analyst" and asks it to rank the victim's assets as high, medium or low value, and to judge their password habits and security.

The default is Anthropic's Claude Opus 4.6. If a model refuses, VOLTA switches to Moonshot AI's Kimi K2, then Google's Gemini 2.5 Flash. If no API key is present, it skips the AI stage and carries on.

On deletion: the source says the main executable may delete itself after the data leaves the machine. It does not say the model decides what gets deleted, so the headline's framing goes further than the text we have.

Why Talos sees this as a trend

Talos calls VOLTA a mature info-stealer with an AI layer added at the back end, with one purpose: cutting the attacker's analysis workload. It also expects that resisting model refusals may become a standard requirement for malware as these tools mature.