SecurityNews

New VOLTA infostealer uses AI to appraise its victims’ value

Cisco Talos has discovered an infostealer that sends stolen data to an AI model in order to rank victims by financial value. The malware, named VOLTA, combines passwords and tokens with order histories and purchase receipts to build a profile of the victim. VOLTA then deletes itself. The malware does not use any infrastructure of the attackers’ own.

Cisco’s threat intelligence team found VOLTA using CAIRN, a research tool the team published in September 2026. CAIRN detects, classifies and tracks malware with integrated AI. According to Talos, the tool analyses metadata only, meaning researchers do not need to download or execute the malware. The technical details are set out in the blog post Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer.

Wide range of stolen data

According to Talos, VOLTA steals saved passwords and cookies from Chromium browsers, payment card details and crypto wallet recovery phrases. The malware also collects session tokens for Discord, Telegram, Steam and Roblox, Microsoft OAuth tokens, SSH and cloud keys, and Wi-Fi profiles.

Using the hijacked sessions, VOLTA searches the victim’s Gmail account for purchase receipts from Amazon, Walmart, PayPal, Uber Eats and DoorDash, among others. On that basis, the malware confirms the victim’s home address. For Uber Eats and DoorDash, VOLTA also retrieves the order history directly via the stolen sessions. Traditional infostealers focus mainly on login credentials and financial data. VOLTA goes further by building a profile of a person’s digital life.

Language model ranks victims

VOLTA bundles the collected data locally into a file and sends it via the OpenRouter service to a language model. By default, this is Anthropic’s Claude Opus 4.6. If that model refuses the request, the malware falls back to Kimi K2 and then to Gemini 2.5 Flash. The model sorts the information into three tiers (low, medium and high) based on how easily it can be monetised. It also assesses the victim’s security habits.

Talos reports that the prompt also asks the model to draw connections between sparse, unrelated pieces of data. One example is linking a victim’s name, address or university to patterns in their passwords. This allows attackers to compile a list of victims ranked by value and approach them in a targeted way using their own data. The model thus takes over work that criminals previously did by hand.

Jan Heijdra, Field CTO Security & AI at Cisco Benelux, says that with VOLTA, AI acts as an analyst for the attacker, lowering the threshold for targeted attacks. He advises security teams to extend their detection with behavioural analysis so that they can also recognise the automated decisions behind an attack. According to Heijdra, the approach is not tied to specific services, and criminals can apply it to platforms widely used in the Netherlands as well.

No infrastructure of its own

VOLTA relies exclusively on third-party services and uses no attacker-controlled servers. According to Talos, this means there is no central infrastructure through which victims can be counted or the attack stopped. How many systems have been infected is unknown. Defenders must rely mainly on traces on the endpoint and on behavioural patterns.

Talos recommendations

An affected system contains the log file volta_debug.txt. This is a timestamped harvest log that records which data was collected and whether the call to the language model succeeded. Talos recommends securing this file before cleaning the system. In addition, revoke all potentially compromised tokens and change passwords.

If a crypto wallet’s recovery phrase was stored in plain text on the disk, Talos considers the wallet fully compromised. The advice is to move the funds to a new wallet with a seed generated on an air-gapped device with no network connection. Talos also recommends setting up application control policies that prevent Chrome from being launched with the ‘–headless’ parameter.

The full report can be read at Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer.

This article was originally published in Dutch on MSP Business. Read the original →