A Windows trojan sorts its victims with AI

Cisco Talos describes VOLTA-AUTODOXER, a Windows infostealer that combines data theft and automated asset ranking of a victim by a language model. The goal is not just to loot, but to prioritize what can bring profit.

Cisco Talos published on October 31, 2026, an analysis of VOLTA-AUTODOXER, a Windows infostealer that steals credentials and then passes the stolen data through a language model to rank a victim’s monetizable assets. Researcher Ryan Fetterman describes malware that no longer just exfiltrates secrets: it also prepares an exploitation dossier, with sorted and justified valuable elements. In this context, AI serves less to speed up the attack than to transform a simple theft into a better-targeted resale file.

When malware acts as an analyst

Talos presents VOLTA as a well-established stealer, with an optional AI integration layer added. The originality lies in the function assigned to the model: to act as an intelligence analyst serving the operator, not just as a technical component of the malware. The system produces a ranking of the victim’s monetizable assets, with supporting citations, from a raw file of credentials, sessions, and metadata.

This logic changes the usual mechanics. Instead of just filling data archives, the attack prioritizes what deserves immediate attention: bank cards, access tokens, email accounts, game sessions, or professional profiles. The operator thus receives an actionable summary without having to sort through it themselves. It’s cleaner. And far less reassuring.

What VOLTA sucks from the compromised machine

One part of VOLTA collects browser credentials, cookies, saved payment cards, but also data from Firefox, Thunderbird, Outlook, Slack, Steam, Battle.net, Discord, Telegram, Roblox, and Microsoft OAuth tokens. The malware also sucks SSH and cloud configuration files, Wi-Fi profiles, screenshots, and, according to Talos, attempts to crack crypto wallets on site to note the result in a dedicated line. The goal is broad: to reconstruct the victim’s digital identity and operational access.

The stealer doesn’t stop at raw copying. It reuses stolen sessions to enrich the file: searching for receipts in Gmail to corroborate an address, checking order history on Uber Eats or DoorDash, verifying Roblox inventory, or extracting clues about Slack spaces. All is then assembled locally into a structured package, with geolocation, employer domain, presumed birth year, last card digits, and clues about iPhone backups. The generated files are named Identity_Report.txt and a JSON export.

OpenRouter, fallback and refusal: the AI chain is designed to persist

Talos states that the bundle is sent to OpenRouter’s chat-completions endpoint with a system prompt asking the model to act as “a forensic intelligence analyst drafting a target assessment.” The expected output notably includes a section WHAT THEY’RE WORTH, ranking assets by monetization value, and a section THEIR OPSEC, assessing the target’s security hygiene. The prompt also pushes the model to link scattered clues, like a name, address, or university, to possible password patterns.

The embedded configuration uses by default anthropic/claude-opus-4.6, then switches to moonshotai/kimi-k2, and finally google/gemini-2.5-flash if needed. The temperature is set to 0.2, which Talos interprets as a choice favoring coherence over creativity. The malware also tests model refusals—for example, the presence of phrases like “I can’t”—before stepping down one level in the chain. If no API key is available, the LLM layer can even be disabled.

A compilation tool sold as a kit

Talos believes VOLTA is not a fixed binary but a model compiled from a template with six slots replaced at build time. This architecture points to a builder distribution mode: an actor sells access, the buyer provides parameters, then a customized executable is generated for each order. If this hypothesis is correct, the AI component becomes a commercial feature, accessible even to less sophisticated operators.

Parameters related to Telegram, API keys, and settings are encrypted with a key derived from two constants present in the binary. Talos considers this protection weak, as someone reversing the file can retrieve the values. This point matters because it shows that AI is not a gimmick slapped onto amateur code: it is integrated into a distribution chain designed to be monetized.

Talos’s recommendations against this type of stealer

Part of the theft relies on bypassing App-Bound Encryption in Chrome, a mechanism that ties certain sensitive data to the application using it. Talos recommends limiting processes allowed to launch Chrome in headless mode to block this decryption path and alerting when an unexpected process launches the browser without an interface. The company also advises hardware storage of credentials, then revocation of compromised tokens and password changes after an incident.

The volta_debug.txt file, which precisely logs what was collected and whether the model call succeeded, should be preserved during investigation. When seed phrases have been left in clear text on disk, Talos considers the wallet fully compromised and funds should be migrated to a new wallet generated on an isolated machine. The underlying angle is clear: AI does not bring spectacular speed here, but shifts the effort onto the victim, who unwittingly funds inference run on their own machine.

By the numbers

  • October 31, 2026 — publication of the analysis by Cisco Talos.
  • 12:46 UTC on October 11, 2026 — first observation of the sample by Talos.
  • 57ef4242666d78c50df145b34023df564f0fad38812d6a580e66433d427e2279 — SHA-256 hash of the analyzed binary.
  • 3 models — Claude, Kimi, then Gemini in the fallback chain.
  • 6 slots — swappable parameters in the builder according to Talos.

To read