Talos finds VOLTA malware uses Claude to rank stolen data
The self-deleting Windows infostealer sends victim data through OpenRouter to an AI model, then falls back to Kimi K2 and Gemini 2.5 Flash if refused.
Cisco Talos researchers have documented a self-deleting Windows infostealer, named VOLTA-AUTODOXER, that sends stolen data to an AI model to produce a ranked "target assessment" of the victim, according to tomshardware.com. The malware requests that assessment through the OpenRouter model router, with Anthropic's Claude Opus 4.6 set as the default model at a temperature of 0.2.
If the model refuses the request or returns a thin reply, VOLTA switches to Moonshot AI's Kimi K2 and then to Google's Gemini 2.5 Flash. The malware assembles an Identity_Report.txt and JSON bundle locally, saves any AI reply as AI_Analysis.txt and exfiltrates both in a volta_dox.zip archive.
The stolen material covers browser passwords, cookies, payment cards, email profiles, Discord tokens, OAuth tokens, SSH and cloud configuration files, and saved Wi-Fi passwords in plaintext. VOLTA also searches the victim's Gmail for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal, and values Roblox items through Rolimons.
The payload carries a template with six swappable slots for a Telegram channel, API keys and settings, filled at build time, which Talos says indicates the malware is made for resale. It bypasses Chrome's App-Bound Encryption by launching Chrome headless, a step an application-control policy restricting headless Chrome can block.
Without an API key
Even without an OpenRouter API key, the malware can still send the dossier over Telegram, minus the AI analysis. Talos advises victims to preserve the volta_debug.txt log and to revoke or rotate all API keys and tokens.
Claude Opus 4.6 is described as an older model released in February.
Quick answers
What is VOLTA-AUTODOXER?
It is a self-deleting Windows infostealer documented by Cisco Talos that sends stolen data to an AI model to produce a ranked target assessment of the victim.
Which AI models does the malware use?
It requests the assessment through OpenRouter with Claude Opus 4.6 as the default, and falls back to Moonshot AI's Kimi K2 and Google's Gemini 2.5 Flash if the model refuses or returns a thin reply.
How can victims respond?
Talos advises preserving the volta_debug.txt log and revoking or rotating all API keys and tokens.