# VOLTA, the thief that asks Claude how much you are worth

> VOLTA steals passwords and seeds and runs the loot through Claude, Gemini and Kimi to rank victims by economic value before selling them.
> By José Maldonado · Cybersecurity · 2026-11-03

A new information stealer for Windows called VOLTA collects passwords, cookies and wallet seeds, like any other stealer, and then runs the loot through a language model to decide which victim deserves attention first. The research by [Cisco Talos](https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/) describes a program that builds a dossier on the person behind the machine and sends it to Claude Opus 4.6, with Kimi K2 and Gemini 2.5 Flash as alternatives if the model refuses to cooperate.

The novelty is not in the theft, it is in the classification. For years, credential markets have been organized by country, by service and by account price. VOLTA adds a layer that until now was handled by a human analyst: estimating how much money is behind a specific person and how much effort is worth spending pursuing them.

That shift has consequences for anyone who uses the same computer for work and for their personal life. A machine like this does not deliver a loose batch of passwords, it delivers a coherent story with a name, address, company, estimated income and a list of services where action can be taken immediately, without the buyer having to investigate anything. For the attacker, the analysis work is already done and costs not a single extra minute.

## A dossier, not a list of keys

The program collects credentials from Chromium-based browsers, data from Outlook profiles and Slack workspaces, payment cards and cryptocurrency wallet seed phrases. It adds Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud provider keys, Wi-Fi profiles and screenshots.

With that material, VOLTA assembles what its analysts call an evidence package on the person, and it does so before the device owner suspects anything. It includes geolocation and home address, employer and corporate domain, inferred year of birth, the last four digits of a card, Slack workspaces, document authorship metadata, traces of iPhone backups and even a valuation of the victim's Roblox account.

The next step is the one that breaks the mold. The program searches the Gmail inbox for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal to confirm the postal address. In other words, it uses the victim's own email to verify the data it has just stolen and complete the profile with cross-checked evidence.

## Three models and a 240-word prompt

Once the package is assembled, the program serializes it and sends it to [OpenRouter](https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/), a service that routes requests to different models. The default configuration points to Claude Opus 4.6. If the response looks like a refusal, the system detects phrases such as "I can't" or "I'm not able to" and tries Kimi K2 and then Gemini 2.5 Flash in sequence, up to three models.

That reserve mechanism is the most revealing part of the design. The malware's author assumed from the outset that the models would resist and built a system to get around that resistance without human intervention. It is not an implementation detail, it is a product decision.

The system prompt is 240 words long and instructs the model to act as a forensic intelligence analyst. The question it is asked is not what the loot contains, but how much it is worth and in what order it should be exploited. A model that answers that turns indiscriminate theft into a prioritized portfolio.

## The stealer as an analytics service

The [malware economy](https://observatorioblockchain.com/en/cybersecurity/yoba-malware-firmas-chatgpt-claude-github/) has been moving toward the subscription model for years. The program is rented out, the logs are sold and the buyer decides what to do with them. VOLTA suggests a further step: the seller delivers the report already classified, with victims ranked by profitability, and charges for that layer of analysis.

For the buyer, the advantage is obvious. They don't need to comb through ten thousand records to find three interesting corporate accounts. They get a short list with context and with the assurance that the data has been verified against the victim's email. The dirty work is already done.

For the victim, the problem is that their personal email becomes a source of evidence against themselves. The home address that no one writes on a form shows up on a food delivery receipt, and that detail makes it possible to go from a loose credential to an impersonation using real data.

## What can be done about this?

The immediate defense remains the usual one, albeit with new nuances. A password manager with two-factor authentication curbs credential theft, and keeping cryptocurrency wallets on isolated hardware prevents a seed stored in the browser from ending up in a case file. None of this is new, but the context makes it more urgent.

There are two measures that do change compared with the classic stealer. The first is to review which email accounts contain identification data and which services retain them. The second is to treat the personal device as part of the corporate attack surface when it is used for work, because the VOLTA package mixes Slack, Outlook and Microsoft tokens with the Roblox account.

### A warning for companies

Companies should read this case as a warning about hybrid work. A corporate laptop with strict controls and a personal computer with the same email and the same Slack together form a single target. The attacker does not distinguish between the secure device and the one that is not, only between the one that holds the data and the one that does not.

The market for initial accesses has operated for years on lists and per-service prices. A corporate email is worth more than a personal one, an account with access to an admin panel is worth more than a streaming subscription, and the buyer usually discovers the real value after paying. Prior classification changes that economy.

If a seller can state precisely which victim has a specific card, a large company behind them and the ability to move money, the price goes up and the buyer saves time. That margin is what funds the sophistication of the program. AI does not make the theft harder, it makes the sale more profitable.

A problem also emerges for insurers. A claim is assessed by the damage caused, and an attacker who chooses better whom to target produces fewer incidents with greater impact. Risk models built on frequency lose accuracy when the attacker prioritizes.

## The next step is no longer malware

The uncomfortable part of VOLTA is that its advantage does not depend on a vulnerability. It does not exploit a Windows flaw or a configuration error; it uses legitimate tools, such as a model routing service or an email API, to do work that previously required a person with judgment. The line between abuse of services and a cyberattack blurs right there.

Model providers can block obvious patterns, and some will. But VOLTA's design anticipates that blocking with a chain of fallbacks, and each new model adds one more door to that chain. The race is not won with a filter, it is won with cost and latency, and the attacker can afford to wait a few seconds.

The question this case leaves behind is not how to detect one more stealer, but what happens when any data-capture program incorporates an analyst. If the value of the loot depends on knowing whom to squeeze, personal information becomes the most profitable asset in the entire criminal ecosystem.
