Cisco’s Talos research organization has identified a new strain of Windows malware that uses artificial intelligence to evaluate the value of stolen victim data. The infostealer, dubbed VOLTA-AUTODOXER, collects personal information and financial details from infected computers, then sends the data to an AI model to generate a ranked assessment of the victim’s assets.
Once the analysis is complete, the malware deletes its main executable to hinder forensic investigation. However, it leaves behind a debug log file that records the theft activity and the results of the AI query, which researchers say can help victims understand the scope of the breach.
VOLTA Malware Data Collection and AI Integration
VOLTA operates by stealing a wide range of data from the victim’s system. This includes browser passwords, cookies, saved payment cards, and email profiles from clients like Thunderbird and Outlook. It also targets Roblox accounts, Microsoft OAuth tokens, screenshots, and configuration files for SSH and cloud providers. Every saved Wi-Fi password is extracted in plaintext.
The malware goes beyond simple data theft by performing a live enrichment phase. It searches through stolen Gmail sessions for receipts from major retailers and food delivery services like Amazon, DoorDash, and Uber Eats. This allows the malware to confirm the victim’s home address and order history. It also reuses stolen Discord tokens to fill out profile details and values Roblox items using external pricing data.
The collected information is assembled into a structured evidence bundle. This bundle includes geolocation data, home addresses, employer details, inferred birth years, and the last four digits of payment cards. The malware then formats this data into a JSON file and sends it to OpenRouter, a model routing service, using Anthropic’s Claude as the default AI model.
The system prompt instructs the AI to act as a forensic intelligence analyst. It asks the model to provide a blunt, specific, and accurate assessment of the victim’s suitability as a target. The AI is tasked with ranking the victim’s valuables by their potential cash value and analyzing their security posture, including password hygiene.
Resilience and Self-Deletion Mechanisms
VOLTA is designed to be resilient against AI refusal mechanisms. If the default Claude model refuses to answer or provides a thin response, the malware automatically switches to alternative models, including Kimi K2 from Moonshot AI or Gemini 2.5 Flash from Google. If no API keys are available for the model router, the malware can still send the dossier over Telegram without the AI analysis.
The malware is built as a product for resale. It uses a template with swappable slots for custom Telegram channels and API keys, which are filled in at build time. This allows attackers to create custom binaries for their specific needs. The credentials in these slots are encrypted, but researchers note they can potentially be recovered from the file.
After infection, the main executable deletes itself to remove traces of the attack. However, a file named volta_debug.txt is preserved. This log shows what data was collected and whether the AI call was successful. Researchers advise victims to revoke all API keys and tokens, change passwords, and move funds from any compromised cryptocurrency wallets to new ones generated on secure devices.
The malware also attempts to bypass Chrome’s App-Bound Encryption by launching the browser in headless mode. Security experts suggest that application-control policies limiting headless Chrome launches can reduce exposure to this specific attack vector, though it does not stop other forms of data theft or exfiltration.
Source: Tom’s Hardware
