Volta malware uses AI to hunt down victims

Technical reports have revealed a new piece of malware named “Volta Autodoxer,” which uses artificial intelligence models from major companies, including Anthropic, Google and China’s Moonshot, to analyze data stolen from users and classify them according to their financial status and how easily their accounts can be breached.

According to a report published by the US website Wired, the malware does not use AI to carry out the breaches or steal the data, but employs it at a later stage to organize the information it collects and turn it into detailed profiles of victims that help hackers choose their targets and estimate the potential financial return from targeting them.

Volta gathers a wide range of data, including passwords and cookies from internet browsers, email accounts and Slack workspaces, payment card data and cryptocurrency wallets, as well as sessions for apps such as Discord, Telegram, Steam and Roblox, two-factor authentication codes and access keys for cloud services and Wi-Fi networks, in addition to screenshots.

The malware does not stop at collecting this information; it links it together and verifies it, taking advantage of access to victims’ Gmail accounts, analyzing the content of incoming messages and extracting additional data that helps build a more detailed picture of the targeted individuals.

It then sends the data to large language models through the API of the OpenRouter platform, asking them to classify victims according to the size of their wealth, the expected financial return from targeting them, and how easily their accounts can be breached, as well as to assess aspects related to their digital security.

Security researcher Talo Ryan Vittermann of Cisco notes that the malware treats AI models as tools for analyzing and sorting information rather than for carrying out attacks, which makes it closer to a commercial system that other hackers could benefit from.

To circumvent the restrictions imposed by AI companies to prevent their models from being used in harmful activities, Volta presents the data to the models in packages that appear to be materials for intelligence or forensic analysis, instead of asking for help in stealing information or carrying out breaches.

The malware relies on a multi-model system through the OpenRouter platform, starting with the “Opus 4.6” model, then moving to “Kimi K2” from China’s Moonshot, and on to “Gemini 2.5 Flash” from Google if one model refuses to process the request. It also scans responses for refusal phrases before transferring the task to another model.

The danger of this malware lies in its ability to turn massive amounts of stolen personal data into organized profiles that link digital information to its owners and their financial and professional reality, making it easier for hackers to identify the most viable targets and the vulnerabilities that can be exploited, even if they lack sufficient technical expertise.

The report warns that the spread of this type of malware, or its transformation into a tool available for sale, could raise the efficiency of hacking and phishing campaigns by enabling less experienced groups to target specific victims more precisely. It also highlights the risks of employing AI models to support criminal activity, even when they do not directly take part in carrying out attacks.