Spooky New Infostealer Malware Uses AI to ‘Autodox’ Victims

Security researchers at Cisco Talos have uncovered an infostealer strain that wires commercial large language models directly into its workflow, using them to automatically profile victims and surface the most lucrative targets. The malware refers to itself internally as "Volta-Autodoxer," and Talos disclosed the finding exclusively to Kernel Panic. The researchers flagged it using CAIRN, a new Talos framework built specifically to identify hacking tools that incorporate AI capabilities, though they noted they have not tracked whether the malware has actually been deployed in the wild.

Volta's traditional collection layer is broad: it sweeps browsers for saved passwords, Microsoft OAuth tokens, cookies, payment cards, and crypto wallet seed phrases, alongside live sessions in Discord, Telegram, Steam, and Roblox — a target set that appears aimed at younger internet users with dense digital footprints. It then runs a second, deeper pass, using a compromised Gmail or other central account to reconstruct the victim's location and habits from receipts and order history in services such as UberEats and DoorDash.

That harvested data is compiled into a serialized dossier and routed through OpenRouter, a service that dispatches prompts to multiple AI providers. The malware defaults to Anthropic's Claude Opus 4.6, but carries fallbacks to the Chinese open-weight Kimi K2 from Moonshot AI and, failing that, Google's Gemini 2.5 Flash — anticipating that a model's guardrails might refuse the request. Attackers also favor this pattern because it lets them rely on cheaper, older-generation models. Talos researcher Ryan Fetterman noted the malware still functions without the AI calls, which he described as "above and beyond."

The creators embedded a 240-word prompt instructing the model to act as "a forensic intelligence analyst drafting a target assessment from a structured evidence bundle," producing an executive summary, an account of what the person does, "what they're worth," and a read on "their opsec," while connecting dots across fields — for instance, checking whether a Wi-Fi password matches a last name. "Go deep, not wide," the prompt directs. Fetterman said the wording is carefully crafted to avoid overtly malicious framing, but that the operators plainly expected refusals and planned a cascading failover to more permissive models.

The context around infostealers raises the stakes: the malware class has been tied to major breaches, including the Snowflake-linked intrusions at AT&T and Ticketmaster and the theft of roughly a terabyte of Disney data, and has yielded hundreds of millions of credentials in recent years. Fetterman said Volta contains clues suggesting it was built to be sold to a range of attackers rather than developed for one actor's campaigns, echoing Crowdstrike's recent observation of four distinct groups using AI tools against targets in government, health care, retail, and travel.

What to watch: whether Volta appears in real incidents, how aggressively providers clamp down on routing that enables this kind of analysis, and how quickly faster AI-assisted triage changes the economics of stolen credentials — tokens expire and passwords get reset, so speed now translates directly into usable access.