VOLTA infostealer uses AI to rank stolen victim data by value
A Windows infostealer called VOLTA-AUTODOXER adds an artificial intelligence analysis stage to credential theft, turning information collected from a compromised device into a ranked assessment of the victim’s potential value to criminals. The CAIRN research effort found that the malware sends a structured dossier to a large language model (LLM), which is asked to identify monetizable assets and assess the victim’s security practices.
The approach shifts AI from a supporting role in malware operations to a form of post-theft analysis. Rather than simply collecting data for an operator to review, VOLTA attempts to connect information from different accounts and services, then prioritize targets and assets. That could help an operator decide where to focus without manually sorting through every stolen record.
From credential collection to victim profiling
VOLTA collects a broad range of information, including browser passwords, cookies, saved payment cards, cryptocurrency wallet data, and account sessions for services such as Discord, Telegram, Steam and Roblox. It also targets Microsoft OAuth tokens, SSH and cloud-provider keys, Wi-Fi profiles, and screenshots. The malware can extract data from several browser and email applications, as well as Slack workspaces and gaming accounts.
Its follow-up enrichment stage uses stolen, authenticated sessions to gather additional details. The observed activity included searching Gmail for purchase receipts from retailers and delivery services to help confirm a home address, querying Uber Eats and DoorDash order histories, and using a Discord token to collect profile information. VOLTA also checks Roblox inventory values through a lookup at Rolimons.com. The resulting data can link a person’s identity, home and workplace details, online accounts, and financial or digital assets.
The stealer assembles this material locally in a file named Identity_Report.txt and serializes it as JSON for submission to OpenRouter. Its prompt directs the model to prepare a target assessment, including a ranked estimate of valuable assets and an evaluation of the victim’s operational security and password hygiene. The model is also asked to support its assessment with citations, making the output more useful as an operator’s triage document.
Model failover and data handling
The observed configuration uses anthropic/claude-opus-4.6 by default, with moonshotai/kimi-k2 and google/gemini-2.5-flash as fallback models. VOLTA checks responses for refusal language and can move to the next model if one declines the request. The malware sets a temperature of 0.2, a relatively low value intended to produce more consistent output. The AI stage can also be skipped if the required key is absent.
After analysis, VOLTA saves the returned material as AI_Analysis.txt. The research describes several possible exfiltration stages: an initial volta_log.zip, a volta_dox.zip containing the AI analysis, and a volta_doxnet.zip when network enrichment yields additional information. The model request runs through the victim’s own internet connection before the result is sent out. Researchers reported no attacker-owned infrastructure in the observed setup; the malware uses third-party services and configurable values supplied when a build is prepared.
VOLTA also includes an embedded cryptocurrency wallet-vault cracking capability and can write recovered results into its own report. Its collection of Wi-Fi credentials and its attempts to access browser-stored secrets expand the consequences beyond the initial theft of a password. The research notes that the malware’s token theft relies in part on techniques to bypass Chrome App-Bound Encryption.
Defensive actions for enterprise teams
Security teams should treat infostealer exposure as a potential compromise of both credentials and active sessions. Revoke potentially affected tokens and sessions, reset passwords, and review identity-provider and cloud audit records for suspicious access. Where an incident involves cryptocurrency seed phrases stored in plaintext, the researchers advise treating the associated wallet as compromised and moving funds to a new wallet with a seed generated on an air-gapped device.
For Chrome environments, the researchers recommend application control that limits which processes can launch Chrome with the –headless option, as well as alerting on launches from processes that are not approved developer tools. These controls address a technique used in VOLTA’s attempt to access browser-protected data. Organizations should also favor hardware-backed credential storage where available and minimize the persistence of reusable tokens on endpoints.
Because the main executable is designed to delete itself, incident responders should preserve the file %APPDATA%\volta_debug.txt if present. It is a timestamped harvest log that can show what data was collected and whether the LLM request succeeded. The sample identified by researchers has SHA-256 hash 57ef4242666d78c50df145b34023df564f0fad38812d6a580e66433d427e2279.
VOLTA illustrates a practical criminal use of generative AI: reducing the human effort needed to interpret stolen data, rather than automating the initial compromise. For defenders, this makes rapid token revocation, endpoint investigation, and careful review of identity and account activity important even when a malware executable is no longer present.
