VOLTA infostealer turns stolen data into AI dossiers
The VOLTA infostealer, Windows malware designed to steal information, is using artificial intelligence to turn raw account data into ranked dossiers about victims and their potential financial value. Cisco Talos says the malware, also called VOLTA-AUTODOXER or VOLTA, submits collected evidence to a large language model (LLM), an AI system that analyses and generates language, so operators do not have to review every stolen record themselves.
VOLTA collects a wide range of personal and business data. Its targets include passwords and cookies, which are browser records that can keep a user signed in, from Chromium-based browsers. It also searches Outlook profiles and Slack workspaces, payment-card details, cryptocurrency wallet seed phrases, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens used to authorise account access, Wi-Fi profiles and screenshots. The malware additionally takes SSH and cloud-provider keys, digital credentials that can permit access to remote systems and cloud resources.
Cisco Talos found that VOLTA assembles this material into a structured evidence bundle intended to describe the person using the compromised computer. The dossier can contain geolocation and home-address information, the victim's employer and employer domain, an inferred birth year, the final four digits of cards, Slack workspaces, document authorship metadata, signs of iPhone backups and a live valuation of a Roblox account. This gives an operator more context than an ordinary list of stolen passwords.
The malware also uses an existing stolen session to search the victim's Gmail account for receipts. It looks for messages associated with Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal, extracting order information that can corroborate a home address. After collection, VOLTA serialises the evidence, meaning it packages the data into a format another system can process, and sends it through OpenRouter, a gateway that routes requests to different AI models.
The default model is Anthropic's Claude Opus 4.6. If that request fails, VOLTA can move to Moonshot AI's Kimi K2 and then Google's Gemini 2.5 Flash. Talos says the malware checks responses for language indicating that a model has refused the request, such as statements that it cannot comply, before trying the next model in a chain of as many as three. This multi-model design is intended to keep the analysis working when one service rejects the task.
A 240-word system prompt, a set of instructions controlling how an AI model responds, tells the model to act as a forensic intelligence analyst. It also orders the system to rely only on the supplied evidence bundle rather than adding facts from its training data. The requested assessment includes a section ranking valuables by their monetisation potential as low, medium or high, plus an evaluation of the victim's operational security and password habits. Talos characterises the model as a triage analyst that converts a credential log into a prioritised dossier with supporting references to the stolen evidence.
Exfiltration, the unauthorised transfer of data out of a device, occurs through Telegram in three stages. Immediately after execution, VOLTA sends browser passwords, cookies, cards, cryptocurrency information, Discord, Steam and Roblox tokens, Wi-Fi passwords and screenshots. A second transfer follows the AI request and contains both the captured information and the model's analysis. A final package is sent after live enrichment, including the Gmail receipts and order histories found through the stolen session.
VOLTA also contains a build-time template with six configurable slots, including the Telegram destination, application programming interface (API) keys used to access online services and other settings. Talos says this design may indicate that the author is advertising a builder kit to customers or operating a bespoke service that creates a separate executable for each order. That status remains uncertain. If it is being sold as commodity malware, the ranking feature would allow customers with little technical sophistication to receive prioritised victim assessments directly through Telegram. Talos describes VOLTA as a mature stealer to which a late-stage AI component has been added chiefly to reduce the analytical work required from its operator.
The findings illustrate a broader use of generative AI in criminal workflows. Zscaler ThreatLabz says the technology is suited to classifying and searching large quantities of material after an attacker gains access to a system. It can help locate valuable or sensitive data for theft or for encryption during a later phase. A ransomware operator, for example, could use this approach to identify the most sensitive information in a compromised network and increase pressure on the victim to pay.
Cisco Talos also recently described NIK C2, a Go-based Windows remote access tool that provides persistent control over compromised hosts. Command-and-control (C2) is the infrastructure through which an attacker sends instructions to infected devices. NIK C2 supports more than 52 commands and includes remote access, credential theft, keylogging, which records keyboard input, surveillance, defence evasion and ransomware capabilities. It also integrates an OpenAI client so stolen data can be analysed during the same compromised session rather than processed later.
NIK C2 first checks whether the attacker supplied an OpenAI API key. If not, it searches the victim's environment variables, system settings used by programs, for the string "OPENAI_API_KEY" and uses any matching credential it finds. This is LLMjacking, the theft of somebody else's AI access to run the attacker's work while the legitimate owner pays for the computing resources. Talos says this is a deliberate operational feature rather than a simple backup, and it can also make it harder to determine who submitted the requests. By combining collection and analysis in one continuous session, the framework may remove processing delays that previously gave defenders time to detect activity.
For businesses assessing this kind of credential and infrastructure exposure, AEU-I provides security-first IT, infrastructure and consulting. VOLTA and NIK C2 show that stolen data is no longer merely collected: AI systems can organise it immediately, identify the most valuable accounts and reduce the effort needed to exploit a compromised person or organisation.