Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarWired→ original

Volta Password-Stealing Malware Tasks Claude, Kimi, and Gemini With Profiling Victims

An infostealer is malware that takes passwords, cookies, and login tokens from an infected computer. The new Volta-Autodoxer, discovered by Cisco Talos, goes further: it sends the loot to Claude Opus 4.6 and asks the model to act as an intelligence analyst and compile a dossier on the victim. If Claude refuses, the malware tries Kimi K2, then Gemini 2.5 Flash. That is why the leak of a single account should be treated as the leak of a person's entire profile.

AI-processed from Wired; edited by Hamidun News
Volta Password-Stealing Malware Tasks Claude, Kimi, and Gemini With Profiling Victims
Source: Wired. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceWired ↗
Editor-in-Chief
Key metric #1A 240-word prompt asks the model to assess how much the victim is "worth" and how well they are protected.
Key metric #2The malware works without the AI component; AI analysis is only an add-on.
Key metric #3Talos did not track whether Volta has been used in real attacks.

Imagine a virus that does not just steal your passwords but also reads your email, works out where you live and how much money you have, and hands its operator a ready-made report: "this person is worth attention." That is how the new infostealer Volta-Autodoxer, discovered by researchers at Cisco Talos, operates. An infostealer is malware that usually hides in fake software and takes logins, passwords, cookies, login tokens, and card data from an infected computer. What is new about Volta is that it sends what it collects to commercial AI models for analysis: Claude, Kimi K2, and Gemini. They compile a dossier on the victim, and the attacker immediately sees whose data is worth the most.

Infostealers have long been a major source of high-profile breaches. The logins they stole were behind the attacks on Snowflake customers, including AT&T and Ticketmaster, as well as the theft of a terabyte of Disney data. In total, such malware has stolen hundreds of millions of credentials.

What Volta-Autodoxer Does

The malware gets its name from what it calls itself inside its code. "Doxing" is the collection of personal information about a person, and here that process is automated. As an infostealer, Volta is fairly mature and takes a lot: passwords from browsers, Microsoft OAuth tokens, cookies, payment cards, cryptocurrency wallet seed phrases, and active sessions on Discord, Telegram, Steam, and Roblox. Judging by its target list, it is aimed at young users with a busy digital life.

Then comes a second, deeper pass. Using what it found in the first step, the malware gets into Gmail or another central account and fills out a portrait of the victim. Receipts and order history from UberEats and DoorDash reveal where the person lives and what their habits are. Everything is assembled into a single structured file that goes to a language model for analysis.

How the Malware Talks to AI Models

Requests are routed through OpenRouter, a service that forwards prompts to various AI providers. By default, it uses Claude Opus 4.6 from Anthropic. The authors apparently expected the model's safety restrictions might block the request, so they added fallback options: Kimi K2, an open-weight Chinese model from Moonshot AI, and, if that also refuses, Gemini 2.5 Flash from Google. Fallback channels and cheaper models from earlier generations are standard practice for attackers.

The 240-word prompt, with bulleted and numbered lists, asks the model to play a "forensic intelligence analyst" and produce a target assessment. It is supposed to write a brief summary of the person, what they do, "how much they are worth," and how careful they are with operational security. It also explicitly asks the model to "connect the dots between fields," for example by checking whether a Wi-Fi password matches the person's surname.

The requirement is simple: "deep, not wide." Talos researcher Ryan Fetterman notes that the text was written so as not to look overtly malicious. The authors clearly prepared for refusals on the first request.

At the same time, the implant still functions without AI, which serves only as an add-on.

Why This Matters

Authentication tokens have a short shelf life, and passwords get changed. The faster an attacker picks the most valuable targets, the more time they have to use what they stole. Until now, that sorting had to be done by hand. Now it can be handed to a model that is available even to novice attackers. Fetterman calls this alarming: advanced capabilities are reaching less experienced attackers. In his view, Volta was probably built to be sold to various customers rather than for a single campaign.

There are also fair caveats. Talos found the malware using its own CAIRN system, which searches for hacking tools with built-in AI, but it did not track whether Volta has been used in real attacks. For now, this is a finding in the code, not a confirmed wave of breaches. Still, the trend is already visible. CrowdStrike recently reported four groups of varying sophistication, speaking Chinese, Spanish, English, Indonesian, and Russian, that used AI tools against dozens of targets, including government agencies, healthcare, retail, and tourism.

The practical takeaways are these. Users should stay away from "cracked" software, where infostealers commonly hide, enable hardware keys or passkeys, and, after suspecting an infection, revoke active sessions rather than only changing passwords. Security teams need to account for the fact that after an infection, a dossier on an employee can appear almost instantly. Model providers will have to find ways to detect prompts that are deliberately written without obvious signs of malicious intent.

Story

Key Questions Answered

What Volta-Autodoxer Does?

The malware gets its name from what it calls itself inside its code. "Doxing" is the collection of personal information about a person, and here that process is automated. As an infostealer, Volta is fairly mature and takes a lot: passwords from browsers, Microsoft OAuth tokens, cookies, payment cards, cryptocurrency wallet seed phrases, and active sessions on Discord, Telegram, Steam, and Roblox. Judging by its target list, it is aimed at young users with a busy digital life. Then comes a second, deeper pass.

How the Malware Talks to AI Models?

Requests are routed through OpenRouter, a service that forwards prompts to various AI providers. By default, it uses Claude Opus 4.6 from Anthropic. The authors apparently expected the model's safety restrictions might block the request, so they added fallback options: Kimi K2, an open-weight Chinese model from Moonshot AI, and, if that also refuses, Gemini 2.5 Flash from Google. Fallback channels and cheaper models from earlier generations are standard practice for attackers.

Why This Matters?

Authentication tokens have a short shelf life, and passwords get changed. The faster an attacker picks the most valuable targets, the more time they have to use what they stole. Until now, that sorting had to be done by hand. Now it can be handed to a model that is available even to novice attackers. Fetterman calls this alarming: advanced capabilities are reaching less experienced attackers. In his view, Volta was probably built to be sold to various customers rather than for a single campaign.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…