What the researchers found
A research team at Cisco Talos has identified a new variant of infostealer malware that integrates artificial intelligence capabilities, according to a report by WIRED in the Kernel Panic newsletter. The program is internally named Volta-Autodoxer and, in addition to classic credential-stealing capabilities, sends the collected data to commercial language models to obtain a structured analysis of the victim. The discovery was disclosed exclusively to the newsletter and identified thanks to CAIRN, a framework Talos developed to recognize attack tools that incorporate AI capabilities.
According to Ryan Fetterman, a Cisco Talos security researcher cited in the material, Volta is a mature infostealer that collects a broad spectrum of information and assembles it into a serialized dossier to be sent to multiple models. The request to the model is to behave like an intelligence analyst and produce an assessment of the victim with supporting citations: a prioritization and analysis function, according to the researcher. Talos has not verified whether the malware has been used in real attacks.
What it collects and how it analyzes it
Volta's traditional functions appear aimed at young users with a broad digital presence. The malware looks for passwords saved in browsers, Microsoft OAuth tokens, cookies, payment cards, and cryptocurrency wallet seed phrases, but also active sessions on Discord, Telegram, Steam, and Roblox. In a second phase, leveraging the initial results, it seeks to reconstruct a more complete picture of the person by going through central accounts such as Gmail and using data such as receipts and order histories from services including UberEats and DoorDash.
The data is then sent to AI tools via OpenRouter, a service that can route requests to different providers. By default, Volta routes requests to Anthropic's Claude Opus 4.6; to deal with possible blocks by the model's protection systems, it provides fallbacks to Moonshot AI's open-weight Kimi K2 model and, if that also fails, to Google's Gemini 2.5 Flash. According to Fetterman, building fallback chains and using inexpensive previous-generation models is a common practice among attackers.
The prompt and the limits of the analysis
The malware includes a prompt of about 240 words, with bulleted and numbered lists, that asks the model to draft an assessment of the target based on a structured evidence package. The instructions require an executive summary of the person, details about their activities, an estimate of their value, and a summary of their operational security practices, inviting the model to link data across different fields, for example by checking whether the Wi-Fi password matches the surname. Fetterman observes that the prompt is carefully worded so as not to appear openly malicious and that the malware works anyway even without the AI calls, which represent an additional layer.
The researcher adds that some clues suggest Volta may have been built to be sold to multiple attackers and not for a single campaign. The stated consequence is that, if high-value targets are identified more quickly, the probability increases that stolen credentials and tokens will be used before they expire or are changed. It remains to be verified how widely the malware has actually spread and how effective the model fallbacks are, elements that the material does not clarify.
