Researchers from Cisco Talos have uncovered a Windows-only self-deleting infostealer that brands itself Volta-Autodoxer. The program harvests a broad spectrum of personal data and then hands the raw material to a large language model (LLM) for automatic valuation, effectively outsourcing the analyst’s job to artificial intelligence.
The binary is not a one-size-fits-all tool; each copy is compiled with a custom Telegram channel, API keys and configuration values supplied by the purchaser. Once the infection runs, the main executable erases itself, leaving only a timestamped debug log that records what was collected and whether the AI request succeeded.
Operational flow of Volta-Autodoxer
During the initial phase, the malware extracts browser passwords, cookies and saved payment cards from Chromium-based browsers, as well as credentials from Firefox and email clients such as Thunderbird and Outlook. It also grabs Microsoft OAuth tokens, Roblox accounts, Discord tokens, SSH keys, cloud-provider configuration files, and every saved Wi-Fi password in plaintext. When a cryptocurrency wallet is found, the implant does more than copy the file – it runs an embedded cracker to attempt inline decryption of the vault.
All of the harvested artifacts are assembled into a structured “evidence bundle” that includes geolocation data, home address, employer domain, an inferred birth year, document authorship metadata and the last four digits of payment cards. The bundle is further enriched by re-using stolen Gmail tokens to scrape receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal, by querying Uber Eats/DoorDash order histories, and by valuing Roblox inventories through the Rolimons marketplace.
Once the JSON-encoded bundle is ready, the implant posts it to OpenRouter, a model-routing service. By default it calls Anthropic’s Claude Opus 4.6 (released in February) with a temperature of 0.2, and a system prompt that tells the model to act as “a forensic intelligence analyst” and to be “blunt, specific, accurate.” The LLM returns a concise report that ranks the victim’s valuables by monetization potential, adds an operational-security assessment of password hygiene, and cites its sources. The original bundle and the AI response are compressed into a volta_dox.zip archive and exfiltrated over the victim’s own network connection.
If Claude refuses or produces an insufficient answer, Volta-Autodoxer automatically falls back to Moonshot AI’s Kimi K2 and, as a last resort, Google’s Gemini 2.5 Flash. Operators who omit an OpenRouter API key simply skip the AI stage and ship the raw dossier straight to their Telegram channel.
Defensive considerations and mitigation
The most innovative theft vector is the use of Chrome’s headless mode to bypass App-Bound Encryption (ABE) and decrypt stored sessions without a graphical interface. Deploying an application-control policy that blocks any process from launching Chrome with the --headless flag effectively neuters this decryption path. Organizations should also monitor for unexpected headless Chrome launches and alert on the flag when observed from unknown binaries.
Because the implant can harvest a wide array of authentication tokens, immediate remediation requires revoking every compromised API key, resetting all passwords and rotating OAuth tokens. Any cryptocurrency wallet whose seed phrase was stored in plaintext must be considered fully breached; funds should be moved to a fresh wallet generated on an air-gapped device.
Forensic investigators should attempt to preserve the %APPDATA%\volta_debug.txt file that the malware leaves behind. It contains a detailed log of the items collected and the success status of the LLM call, which can be valuable for attribution and threat-intel sharing.
Finally, the availability of Volta-Autodoxer as a “builder” service—where buyers receive a uniquely compiled binary for a fee—highlights a growing business model in cybercrime: providing turnkey tools that embed AI capabilities so that even low-skill operators can receive intelligence-level target assessments without manual analysis.
