Cisco Talos finds VOLTA infostealer uses AI to rank victims

Cisco Talos has analysed a Windows infostealer named VOLTA (self-styled VOLTA-AUTODOXER) that not only harvests a broad set of credentials and artefacts but also uses an AI model to prioritise victims. The malware self-deletes after running and avoids direct attacker control of command-and-control, instead routing data via third-party systems.

The operator cannily relies on an AI stage hosted externally, using a large language model to generate a “WHAT THEY’RE WORTH” ranking and an “OPSEC” assessment from the harvested data. Evidence of the workflow was documented by Talos, with the VOLTA framework surfacing through the CAIRN research effort.

VOLTA’s data-gathering is comprehensive: browser passwords, cookies and saved card details from Chromium, plus data from Firefox, Thunderbird and Outlook; Slack workspace data and login sessions from Discord, Telegram, Steam, Battle[.]net and Roblox; Microsoft OAuth tokens; SSH keys; cloud-provider credentials; Wi‑Fi profiles; and screenshots.

The malware also attempts to crack crypto wallet passwords on the host and compiles a detailed victim profile including location, employer, domain, approximate birth year and even the last four digits of payment cards, tapping into Gmail receipts, order histories and Discord metadata to augment its dossier.

The exfiltrated content is packaged into volta_log.zip, volta_dox.zip and volta_doxnet.zip for transmission via Telegram, with the AI analysis stored in AI_Analysis.txt and Identity_Report.txt on the victim’s machine.

Defence and response guidance, drawn from the Talos report, include an application-control policy to restrict Chrome launches with the –headless flag to block App-Bound Encryption bypass, and the use of hardware-backed credential stores. If compromised, security teams are advised to preserve volta_debug.txt, revoke tokens, and treat a plaintext wallet seed as fully exposed.

The AI component can default to Anthropic’s Claude Opus 4.6, with fallbacks to Moonshot AI’s Kimi K2 and Google’s Gemini 2.5 Flash if necessary; the model operates on the victim’s device before data are transmitted.