A Windows information-stealing malware called VOLTA-AUTODOXER, or VOLTA, uses AI models to turn stolen credentials and personal details into ranked assessments of victims’ financial value. The Hacker News reported on November 2 that Cisco Talos researchers discovered the malware, which sends both stolen information and AI-generated assessments to its operator through Telegram.
The collection spans personal accounts and workplace access, putting browser sessions, communications, financial information and infrastructure credentials within its scope. Talos describes the AI component as an analytical layer added to an established stealer: it helps operators decide which stolen assets deserve attention.
From stolen credentials to a personal dossier
VOLTA harvests passwords and cookies from Chromium-based browsers, Outlook profile and Slack workspace data, payment cards, cryptocurrency wallet seed phrases, Microsoft OAuth tokens, SSH and cloud-provider keys, Wi-Fi profiles and screenshots. It also collects Discord, Telegram, Steam and Roblox sessions. Cookies and session tokens can provide access to already authenticated accounts; wallet seed phrases are used to recover cryptocurrency wallets.
Beyond collecting credentials, the malware assembles a structured profile of the person using the machine. Talos researcher Ryan Fetterman said this includes geolocation and home address, employer and employer domain, inferred birth year, card last-four digits, Slack workspaces, document authorship metadata, iPhone backup indicators and a live Roblox account valuation.
VOLTA also uses stolen session access to search the victim’s Gmail account for receipts that can corroborate a home address. For example, it extracts messages associated with Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal to add purchase-related evidence to the profile.
The malware converts the collected bundle into a format suitable for submission to an AI service. Requests go through OpenRouter, a gateway that routes requests to large language models, and default to Anthropic’s Claude Opus 4.6. The fallback models are Moonshot AI’s Kimi K2 and Google’s Gemini 2.5 Flash.
Talos found that VOLTA checks model responses for refusal phrases such as “I can’t” and “I’m not able,” then moves through a chain of up to three models if necessary.
A 240-word system prompt casts the model as a forensic intelligence analyst and directs it to use only the evidence bundle for facts about the subject. The requested output includes a “WHAT THEY’RE WORTH” section ranking assets by monetization potential with low, medium or high tiers. Another section, “THEIR OPSEC,” assesses security practices and password hygiene.
Talos characterizes the result as a prioritized, citation-backed dossier. That describes the researchers’ assessment of the feature; the model’s role is to interpret information the malware has already collected.
Three deliveries to Telegram
VOLTA sends data to its operator at three stages. Immediately after execution, it transmits browser passwords, cookies, payment cards, cryptocurrency wallet information, Discord, Steam and Roblox tokens, Wi-Fi passwords and screenshots. A second delivery follows the AI request and includes captured data alongside the model’s analysis. A third follows live enrichment and contains Gmail receipts and order histories.
The malware uses a template with six slots populated when a binary is built, including the Telegram destination, API keys and other settings. Talos said this suggests the author may be offering a builder kit to customers or compiling custom binaries for individual orders.
If VOLTA is sold as a commodity, Talos argues, its dossier feature could give customers with limited analytical expertise ranked target assessments directly in Telegram. The researchers describe the benefit as reducing the operator’s work, rather than improving speed or scale.
AI analysis can also use a victim’s account
The report places VOLTA alongside broader attacker interest in using generative AI to search stolen information. Zscaler ThreatLabz said such systems can classify large collections and locate valuable or sensitive material accessible from a compromised machine. A ransomware operator could use that approach to select sensitive data for theft and increase pressure on a victim to pay.
In a separate disclosure the previous week, Talos described NIK C2, a Go-based Windows remote-access framework supporting persistent control and more than 52 commands, with credential theft, keylogging, defense evasion, surveillance and ransomware capabilities.
NIK C2 integrates an OpenAI client to analyze stolen data during a compromised session. If the operator supplies no API key, the tool searches the victim’s environment variables for OPENAI_API_KEY and uses it instead, Talos found. This is a form of LLMjacking—using another party’s AI access for an attacker’s operations—which can obscure who initiated requests. For a victim with that key configured, the practical consequence is direct: their own OpenAI account can fund the attacker’s analysis of their stolen data.
