Volta uses AI to rank victims from stolen data

A new information stealer called Volta collects passwords, authentication tokens, payment details and clues from victims’ accounts, then sends the haul to language models for analysis. Cisco Talos found the malware using CAIRN, a framework for detecting hacking tools with AI capabilities. The researchers have not tracked whether Volta has been used in real attacks. Its design points to a shift in what an infostealer can do: not just take credentials, but help an attacker decide whose credentials are worth using first.

From credentials to a profile

Volta searches browsers for saved passwords, Microsoft OAuth tokens, cookies, payment-card data and cryptocurrency wallet seed phrases. It can also access active sessions for Discord, Telegram, Steam and Roblox.

The malware then uses its initial findings to dig further into a victim’s Gmail or another central account. It collects details about the person, their location and habits, including receipts and order histories from services such as UberEats and DoorDash.

Stealers have long been a source of stolen credentials at scale. In recent years, they have been linked to major breaches affecting AT&T and Ticketmaster in attacks related to Snowflake, and to the theft of a terabyte of Disney data. Hundreds of millions of credentials have been stolen with their help.

Volta’s distinction is what happens after collection. Cisco Talos researcher Ryan Fetterman said the malware assembles its findings into a structured dossier and sends it to multiple language models, which are asked to act as intelligence analysts and produce a sourced assessment of the victim, including a target-priority ranking.

The model is part of the toolkit

Volta routes its requests through OpenRouter, a service that lets users send prompts to different AI providers. Its default is Anthropic’s Claude Opus 4.6. If that model’s restrictions prevent it from processing a request, Volta can fall back to Moonshot AI’s open model Kimi K2, then to Google’s Gemini 2.5 Flash.

The 240-word prompt asks for a short profile of the person, their activities, their “value,” and an overview of their digital security. It also instructs the model to connect clues across sources, such as checking whether a Wi-Fi password matches the victim’s surname. The prompt urges deeper analysis rather than wider coverage.

Fetterman said the wording is designed not to state an explicitly criminal purpose. The malware’s creators also appear to anticipate refusals, with a sequence of models that have less restrictive safeguards. The AI features are an add-on: Volta continues to work without them.

Fallback models and older, cheaper options are already common among attackers. The more notable feature here is the handoff: a conventional data stealer gathers the material, while models are asked to turn it into a judgment about the person behind the account.

A shortcut to choosing targets

Researchers are finding more cases of attackers using language models to write code or analyze large collections of stolen data, even as some cybercriminals complain that AI-generated material is flooding criminal forums with low-quality content.

CrowdStrike researchers recently told WIRED about four hacking groups with different skill levels. The groups communicated in Chinese, Spanish, English, Indonesian and Russian, and used AI tools in attacks on dozens of targets worldwide, including government agencies and companies in healthcare, retail and travel.

Fetterman said some details of Volta suggest it may have been built for sale to different attackers rather than for one group’s campaigns. My read is that the pitch is less “AI-powered malware” than a way to make ordinary stolen credentials more actionable. Faster prioritization could matter most in the short window before tokens expire or victims change their passwords—and could make that process accessible to attackers with less expertise.