VOLTA Infostealer Uses AI to Rank Victims by Financial Value
A new Windows malware steals your data and uses LLMs to rank your net worth, turning raw logs into prioritized hit lists for hackers.
Research · Source: Cisco Talos
What happened
Cisco Talos just exposed VOLTA-AUTODOXER. It is a self-deleting Windows infostealer that does more than scrape passwords. After stealing credentials, crypto wallets, and session tokens, it uses a large language model to act as an intelligence analyst. The AI ranks the victim's stolen assets by monetization value.
The malware actively enriches its stolen data before calling the AI. It uses stolen Gmail tokens to read Amazon and Uber Eats receipts to confirm home addresses. It checks Roblox inventory values on Rolimons.com. Then it bundles this data into a JSON file and sends it to OpenRouter. The prompt demands a target assessment with sections titled WHAT THEY'RE WORTH and THEIR OPSEC.
VOLTA anticipates AI guardrails. It defaults to Claude Opus 4.6 but checks for refusal phrases like I can't. If blocked, it falls back to Kimi-k2, then Gemini 2.5 Flash. The entire AI analysis runs locally on the victim's network before exfiltrating the final dossier. Talos notes this is currently a capability report and has not yet been observed in real-world attacks.
Key facts
- 0.2 — The low temperature setting used to ensure consistent, deterministic AI outputs.
- 3 — The maximum number of models in VOLTA's fallback chain to bypass refusals.
- anthropic/claude-opus-4.6 — The default model hardcoded into the malware's OpenRouter configuration.
Why it matters
The bottleneck in cybercrime used to be human attention. Hackers would steal massive databases but lack the time to sift through raw logs to find high-value targets. VOLTA changes this by automating the triage phase. Unskilled buyers can now purchase a custom malware build, deploy it, and receive a prioritized dossier directly in their Telegram inbox.
This shifts the defensive landscape. AI is no longer just writing phishing emails or generating code. It is actively making connections between sparse data points, like linking a victim's university to their password patterns. If malware can stretch scraped data to infer net worth and physical locations, the blast radius of a simple browser compromise expands exponentially.
For builders
Guardrails are a speed bump
VOLTA proves that relying on a single model's safety filters is useless against determined adversaries. The malware simply cascades from Claude to Kimi to Gemini until it gets an answer. Builders must assume attackers will always find a permissive model.
Local inference hides API calls
The malware makes its OpenRouter API calls directly from the victim's machine using their own network connection. This bypasses traditional network defense tools looking for suspicious outbound traffic. Security tools must adapt to monitor anomalous API requests originating from legitimate user endpoints.
AI is now a MaaS feature
VOLTA uses a builder model with swappable slots for API keys and Telegram channels. AI analysis is now a commercial feature sold to low-skill criminals. Security vendors will lose ground if they do not automate their own threat intelligence at the same pace.
My take
The scariest part of VOLTA is not the theft. It is the sorting. AI is doing the grunt work of an intelligence analyst, handing unskilled attackers a ready-made hit list. We are entering an era where you get hacked not because you were targeted, but because an LLM decided your Uber Eats history made you a profitable mark.
Original reporting: Cisco Talos. This is my rewrite and opinion.