Volta infostealer has Claude, Kimi and Gemini profile its victims

What it's about

According to a report shared exclusively with Wired ("Kernel Panic"), researchers at Cisco Talos have discovered a new variant of infostealer malware that builds in a large language model (LLM). Infostealers are malicious programs that steal passwords, authentication tokens, cookies, credit card data and more from infected computers. The new software calls itself "VOLTA-AUTODOXER" internally. It automatically builds profiles of its victims ("doxing" means collecting and combining personal data) and ranks which targets are the most lucrative. It was found using Talos's new CAIRN framework, which is designed to detect hacking tools with AI features. According to Talos, a sample was first seen on October 11, 2026.

What Volta steals

As a stealer, Volta is considered mature by Talos. It grabs browser passwords, cookies and saved payment cards from Chromium browsers, as well as from Firefox, Thunderbird, Outlook, Slack, Steam, Battle.net, Discord, Telegram, Roblox, Microsoft OAuth tokens, SSH and cloud keys, and all saved Wi-Fi passwords. Crypto wallets aren't just copied; a built-in cracker tries to break open the vaults. According to Wired, it appears to target younger users with an intense digital life.

A second round follows. Using stolen sessions, Volta searches the Gmail inbox for purchase receipts (Amazon, DoorDash, Uber Eats, Grubhub, Walmart, PayPal) to confirm the home address, pulls order histories from Uber Eats and DoorDash, enriches Discord profiles and values Roblox inventories through a live lookup. The dossier includes location, address, employer, estimated birth year, the last four card digits and Slack workspaces, among other things.

The AI as analyst

The dossier is created locally as "Identity_Report.txt", packaged as JSON and sent to a model via OpenRouter, a service that routes requests to various AI providers. The request goes out over the victim's own internet connection, from the victim's machine. The reply is saved as "AI_Analysis.txt", and both files go to the attackers together in "volta_dox.zip".

According to Wired, the prompt (about 240 words) opens with "You are a forensic intelligence analyst producing a target assessment from a structured evidence package". It asks for a summary, personal details, a "WHAT THEY'RE WORTH" section rating assets by monetizability (low/medium/high), and "THEIR OPSEC", an assessment of security behavior and password hygiene. The model is told to "connect dots across fields", for example checking whether a Wi-Fi password matches the person's surname. The instruction: "Go deep, not wide." The model isn't supposed to find anything new, but to draw conclusions from the data already there.

Volta checks replies for refusal phrases such as "I can't" or "I'm not able" and then queries the next model. Ryan Fetterman of Talos says the prompt is carefully worded so it doesn't look overtly malicious, and that the developers expected refusals even on the first request. Fallbacks and older, cheaper models are common among attackers, he says.

A product for sale, not a one-off tool

According to Talos, Volta uses a template with six swappable fields (including the Telegram channel and API key) that are filled in at build time. That points to a builder model: the author sells access, customers supply their own values, and a separate binary is generated for each order. The keys are only weakly obfuscated. All observed endpoints belong to third parties; there is no dedicated attacker infrastructure. The malware deletes itself. This means even less skilled customers could get victim rankings delivered straight to their Telegram inbox.

Countermeasures

Much of it relies on bypassing Chrome's "App-Bound Encryption". Talos recommends an application policy that blocks unknown processes from launching Chrome with --headless, or at least alerts on it, plus hardware-backed storage for credentials. If you're infected, preserve the file %APPDATA%\volta_debug.txt: it logs what was collected and whether the AI call succeeded. Revoke tokens and change passwords. If a wallet seed sits in plaintext on the disk, treat the wallet as compromised and move funds to a new wallet.

Context

Talos speaks of "effort displacement": the AI replaces the step that needs the most human attention, picking valuable victims, and makes malware not faster but more convenient. Talos saw the same pattern with CLOSEDQUORUM. A second trend is resilience against model refusals. Wired points to CrowdStrike, which has observed four hacking groups of varying skill (Chinese, Spanish/English-speaking, Indonesian, Russian) using AI to attack dozens of targets. It remains unclear whether Volta is actually in use, since Talos didn't track that. The sources also don't say whether providers can detect abuse through OpenRouter. That fallbacks are meant to circumvent the models' own safeguards is a warning sign for providers.