Volta-Autodoxer infostealer sends stolen data to AI models to profile its victims

Researchers at Cisco Talos have found a strain of infostealer malware that passes the data it steals to commercial large language models and asks them to write an assessment of each victim. The malware calls itself "Volta-Autodoxer" internally. Wired reported the research on October 31, 2026.

The stealing works the way infostealers have worked for years. What is new is the step after it. Sorting stolen data used to be manual work for attackers. Volta hands that step to a model, which ranks victims and pulls out the details an attacker would use first.

What the malware collects

Infostealers are often hidden in counterfeit software. They take usernames, passwords, authentication tokens, cookies and card data from infected machines. Wired notes that many large breaches in recent years started with credentials gathered this way, including the Snowflake-linked breaches that hit AT&T and Ticketmaster.

Volta's first sweep collects passwords saved in browsers, Microsoft OAuth tokens, cookies, payment cards and crypto wallet seed phrases. It also takes active Discord, Telegram, Steam and Roblox sessions. According to Wired, that mix suggests it is aimed at younger internet users with busy digital lives.

A second sweep builds on the first. Volta uses access to a central account such as Gmail to work out where the victim lives and what their habits are. One source it uses is receipts and order history from services such as UberEats and DoorDash.

"Volta is pretty mature as a stealer. It casts a pretty wide net for what it collects," Ryan Fetterman, a security researcher at Cisco Talos, told Wired.

How the AI step works

Volta turns everything it collects into one serialized dossier and sends it to language models through OpenRouter, a service that routes prompts to several AI providers. The default model is Anthropic's Claude Opus 4.6. If that request fails, perhaps because a guardrail blocks it, Volta falls back to Moonshot AI's open-weight Kimi K2. If that also fails, it tries Google's Gemini 2.5 Flash. Wired notes that fallbacks and cheaper models are common attacker techniques.

The instructions to the model are a 240-word prompt. "You are a forensic intelligence analyst drafting a target assessment from a structured evidence bundle," it begins. It asks for an "executive summary" of the person, what they do, "what they're worth" and a summary of "their opsec." It tells the model to "connect dots across fields", for example by checking whether a Wi-Fi password matches the victim's last name. "Go deep, not wide," it says.

Fetterman said the result is "structured citation-backed assessment of the victim," and called the step "a prioritization and analysis function." He said the prompt is "pretty carefully crafted not to imply anything overtly malicious," but that its authors still expect refusals and planned the fallback chain toward "more permissible models." The AI step is optional. According to Fetterman, the malware still works without it.

Talos flagged Volta with CAIRN, its new framework for finding hacking tools that use AI. Fetterman said Talos did not track whether Volta has been used in the wild. He said the code contains clues that it may have been built for sale to many attackers rather than for one group's campaigns.

Part of a wider pattern

Volta is not the only case of attackers putting language models to work. Wired cites researchers at CrowdStrike who said they had seen four separate hacking groups, operating in Chinese, Spanish and English, Indonesian and Russian, using AI tools to compromise dozens of targets. Those targets included government bodies and businesses in health care, retail and travel. Security researchers are also finding attackers using models to write code and to go through large stores of stolen data.

Fetterman explained why speed matters. Authentication tokens expire and passwords get changed. "If you're able to identify your high-value targets faster, you probably have a better chance of using their credentials or authenticating using stolen tokens more often than maybe you would have if you had to do that triage yourself," he said. "We're seeing more advanced capabilities accessible to lower-level sophistication attackers."

What it means for work

Volta automates the analyst work on the attacker's side: reading through stolen logs, linking accounts and deciding which victim to act on first. That makes the time between infection and misuse more important to defenders. Revoking sessions, expiring tokens quickly and rotating credentials all work best before an automated profile turns stolen data into a decision. For individuals, ordinary records such as delivery receipts and gaming sessions now feed a ready-made profile of where they live and what they are worth, not just a list of passwords. The dependence on commercial models gives AI providers and routing services one more place to detect abuse, though the fallback chain shows attackers expect some models to refuse. It is still unknown how widely Volta has been deployed and whether profiling actually leads to more successful account takeovers.