A new Windows infostealer is using an AI model to decide which stolen assets are most likely to produce money. The malware, named VOLTA-AUTODOXER, turns browser credentials, payment data, cloud keys and other stolen material into a ranked victim profile.
The approach matters because it removes much of the manual analysis usually required after malware grabs everything it can find. The victim is no longer just another infected computer. VOLTA creates an evidence bundle, sends it to an AI model and asks for a commercial assessment of the target.
How VOLTA turns theft into a target profile
VOLTA starts by taking a wide range of sensitive material. Cisco Talos researcher Ryan Fetterman says it captures credentials and cookies stored in web browsers, card records saved there, recovery strings for digital wallets, Microsoft OAuth tokens, SSH keys, cloud credentials, screenshots, and active Discord, Telegram, Steam and Roblox sessions.
It also reveals passwords for wireless networks saved on the computer in clear text. A built-in cracking tool targets cryptocurrency wallet files. This indiscriminate collection gives the operator flexibility: the malware does not need to know in advance which stolen item will be worth the most.
The more consequential stage comes afterward. VOLTA assembles what Fetterman described as a structured evidence bundle containing geolocation, home address, employer and employer domain, an inferred birth year, card last-four digits, Slack workspaces, document-authorship metadata, iPhone backup indicators and a live Roblox account valuation.
Stolen session access lets the malware go further. It can search Gmail for purchase receipts to corroborate a home address, pull UberEats and DoorDash order history and enrich Discord accounts. The output is therefore more than a stolen password database. It is a profile intended to help an attacker decide where to focus.
The AI layer changes the economics
After collection, VOLTA creates a local file called `Identity_Report.txt`. The report is serialized as JSON and posted to the OpenRouter API with instructions telling the model to act as a forensic intelligence analyst drafting a target assessment.
The prompt asks the model to rank valuables by how easily they could become money, classify each as low, medium or high, assess security habits and password hygiene, and identify links among scattered details. The returned analysis is saved as `AI_Analysis.txt`, and both files are exfiltrated together in `volta_dox.zip`.
By default, VOLTA requests Anthropic’s Claude Opus 4.6 through OpenRouter. If that model declines, the malware tries Moonshot AI’s Kimi K2 and then Google’s Gemini 2.5 Flash. Fetterman said this fallback design suggests the operator expects refusals and wants the analysis job to continue through smaller, more permissive models.
The operator receives the ranked assessment through Telegram. That is the practical payoff. Instead of reviewing a large pile of stolen data manually, a buyer can begin with a short list of assets and their apparent monetary value.
The configuration also points toward a builder model. Fetterman said the malware author appears to sell access, collect values from customers and compile a custom binary for each order. Settings such as the Telegram channel and API keys occupy configurable slots, although the embedded credentials are not meaningfully encrypted.
Why this matters
Researchers describe VOLTA as a mature stealer with an AI layer added to reduce the analysis burden on the operator. AI does not create the theft; it improves the sorting.
That distinction has operational value for criminals. A broad collection is noisy, while human review takes time. A model can connect a name, address, employer, purchase history, card details and password patterns, then direct attention toward the assets that appear easiest to monetize.
The losers are not only the people whose credentials were taken. Stolen cloud keys, OAuth tokens and active sessions may give an attacker access beyond the infected endpoint. Even material that is not immediately abused can contribute to a more complete picture of the victim.
Password replacement alone does not address every exposed route. Session theft may work without the underlying password, and stolen keys or tokens may remain usable until they are revoked or rotated.
Where defenders should focus
| VOLTA stage | What is exposed or performed | Defender implication |
|---|---|---|
| Collection | Browser credentials, session cookies, payment-card records, cryptocurrency material, OAuth tokens, SSH keys, cloud credentials and service sessions | Treat each credential type separately; a password change does not invalidate every reusable secret |
| Enrichment | Location, workplace, purchase history, account details and document metadata are assembled | Review account activity and recovery settings, not only the malware’s collection list |
| AI analysis | Assets are ranked by their apparent monetary value and assigned low, medium or high labels | Prioritize investigation and credential revocation by actual access and exposure, rather than assuming every stolen item carries equal risk |
| Exfiltration | `Identity_Report.txt` and `AI_Analysis.txt` are packaged in `volta_dox.zip` | Preserve relevant endpoint and telemetry evidence while the exposed credentials and sessions are revoked |
What to watch next
Talos saw the same broader pattern in the CLOSEDQUORUM implant, where AI was applied to a different phase of an attack. Fetterman said the important question is not merely whether the pattern will appear elsewhere, but which attack phase it reaches next.
For VOLTA specifically, watch for expansion of the builder infrastructure, changes to the data collected and additional infostealers adding target scoring before stolen material reaches an operator. The meaningful shift is from “collect everything” to “collect everything, then rank the haul.”
That makes cleanup more urgent. Changing a password may be a start, but revoking active sessions, rotating keys and reviewing account changes are the steps that close routes an attacker can reuse.
Key takeaways
- VOLTA uses AI to rank stolen assets by their apparent monetary value and assess the target’s security habits.
- Password changes do not invalidate every stolen browser session, OAuth token or access key.
- Security teams should treat session revocation, key rotation and account-history review as separate response tasks.
- The broader pattern is to identify the most labor-intensive stage of an attack and replace it with a model call.
Frequently asked questions
What is VOLTA-AUTODOXER?
It is a Windows infostealer discovered by Cisco Talos that collects sensitive data and uses an AI model to rank potential assets by how easily they could be turned into money.
Does VOLTA send stolen data directly to Anthropic?
The malware sends a JSON bundle and prompts to OpenRouter. By default it requests Claude Opus 4.6, with fallbacks to Kimi K2 and Gemini 2.5 Flash if a model refuses.
Why should changing a password not be the only response?
Stolen session cookies, OAuth tokens and other access credentials may continue working after a password changes. Revoking sessions, rotating keys and reviewing account activity address those additional risks.
What makes VOLTA different from a conventional infostealer?
VOLTA still performs broad collection, but it adds an AI analysis stage that turns the stolen information into a ranked target assessment and delivers that assessment to the operator through Telegram.