VOLTA infostealer uses Claude and Gemini to rank stolen data

Cybersecurity researchers have identified a new Windows information stealer, called VOLTA-AUTODOXER or VOLTA, that uses commercial large language models to rank a victim's stolen data by its monetisation value, according to a report published on 2 November 2026 by The Hacker News. The models named in the report are Anthropic's Claude, Google's Gemini and Kimi, the chatbot built by Chinese company Moonshot AI.

According to the researchers, VOLTA behaves like other stealer families in most respects. It harvests saved passwords and cookies from Chromium-based web browsers, which covers Chrome, Edge and other browsers built on the same engine. It also collects data from Microsoft Outlook profiles and from Slack workspaces, giving an operator access to both personal credentials and workplace communications.

The distinguishing feature is what happens after the theft. Rather than dumping everything into a log file for a human to sift through, VOLTA is described as passing the stolen material to a large language model and asking it to score the victim by financial value. In effect, the malware outsources triage, the step where criminals decide which of thousands of compromised machines is worth pursuing, to a commercial AI service. That is a misuse of ordinary model access rather than an exploit of any flaw in the models themselves.

The report does not say how VOLTA is distributed, how many victims have been affected, or which regions or industries have been hit. It does not state whether the malware uses paid developer access to the three services, stolen API keys, or some other route. None of Anthropic, Google or Moonshot AI is reported to have commented, and the sources give no indication that any of the three has suspended accounts linked to the campaign or shipped a change in response.

For defenders, the practical point is narrow but worth noting. Infostealers have long produced far more data than their operators can read. A model that sorts victims by apparent wealth or by the value of the corporate accounts found on a machine shortens the gap between an initial infection and a targeted follow-up, such as account takeover or extortion. The researchers present this as the purpose of the feature; they do not quantify how well it works, and no evidence of successful downstream fraud is described in the material available.

It also means that outbound traffic from an infected machine may look like ordinary AI usage. Requests to well-known model providers are now routine on corporate networks, so traffic to those domains is unlikely to stand out in logs on its own. Organisations that already monitor for stealer activity should continue to focus on the behaviours the report describes directly, namely bulk access to browser credential stores, Outlook profile data and Slack tokens.

Until fuller technical detail is published, the scale of the campaign is unclear. The reported discovery date is 2 November 2026, and the status is a researcher report rather than a confirmed breach at any named organisation. Readers should treat claims about VOLTA's reach, including any estimate of victim numbers, as unverified unless the researchers release indicators of compromise and telemetry to support them.