Cisco Talos has described a new Windows infostealer, VOLTA-AUTODOXER, that feeds stolen victim data to a large language model and asks it to rank what the victim is worth. The model access runs through OpenRouter, with Anthropic’s Claude Opus 4.6 as the default.
What VOLTA-AUTODOXER Steals From Windows Machines
Talos researcher Ryan Fetterman described a broad collection routine. VOLTA steals Chromium passwords and cookies, Outlook and Slack data, payment cards and crypto wallet seed phrases. It also takes Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud keys, Wi-Fi profiles and screenshots.
The Evidence Bundle and Gmail Receipt Searches
The malware does not stop at collection. It builds an “evidence bundle” on each victim containing geolocation, home address, employer, inferred birth year, card last four digits, Slack workspaces, iPhone backup indicators and Roblox account value.
To confirm a home address, VOLTA uses stolen sessions to search the victim’s Gmail for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal.
How VOLTA Sends Victim Data to Claude, Kimi and Gemini
The evidence bundle goes to an LLM through OpenRouter. The default model is Claude Opus 4.6, with fallbacks to Moonshot AI’s Kimi K2 and Google’s Gemini 2.5 Flash. If a model declines the request, VOLTA retries when it sees refusal phrases such as “I can’t.”
The 240-Word Prompt Casting the Model as a Forensic Analyst
A system prompt of about 240 words casts the model as a “forensic intelligence analyst.” It requests a report section titled “WHAT THEY’RE WORTH” that ranks the victim’s valuables by monetization tier. The output turns a pile of stolen files into a prioritized list an operator can act on without reading each item.
Automated Triage Targets Doxxing and Extortion Value
The design shifts a task that once needed a human reading through logs. Stolen data is triaged automatically for doxxing and extortion value, so operators can focus on the victims whose accounts, wallets or personal details promise the largest return. The retry logic against refusals shows the authors expect the models’ safeguards to push back and built a workaround into the malware itself.
Talos has provided detection guidance. The available reporting does not describe how VOLTA is delivered, does not give dates for the campaign and does not disclose victim counts, so the scale of infections is unknown. The Talos report was published on November 2.
The design follows a consistent logic. Collection captures sessions and tokens, enrichment turns them into a profile, and the model produces the ranking. Roblox account value and iPhone backup indicators in the evidence bundle show the author cares about consumer victims whose accounts can be resold, not only business users. Microsoft OAuth tokens and SSH and cloud keys, on the other hand, point to access that could be monetized in other ways.
Talos named the model fallbacks in order: Claude Opus 4.6 first, then Kimi K2, then Gemini 2.5 Flash. A chain of three providers keeps the analysis running if one blocks the request.
The “WHAT THEY’RE WORTH” section and its monetization tiers describe the output the operator receives, and the report treats that ranking as the feature that sets the malware apart from conventional stealers that only exfiltrate files. The reliance on commercial model APIs also means each analysis requires outbound traffic to OpenRouter, a possible observation point for defenders.
The breadth of the collection list shows how many account types sit behind a single infection. A Windows machine that holds Chromium sessions, Slack workspaces, cloud keys and a wallet seed phrase hands the operator several separate routes to monetization, and the model’s ranking tells the operator which route pays best. Talos’s detection guidance is the available defensive reference, and the report gives no further victim figures.
