VOLTA Infostealer Offloads Target Prioritization to Multi-Model LLM Chains
The VOLTA-AUTODOXER malware executes post-compromise asset triage by piping stolen session tokens and local enrichment telemetry into an LLM failover chain via OpenRouter. Operating without attacker-owned C2 infrastructure, the self-deleting binary ranks victims by asset value directly on the host prior to exfiltration.
Host Harvesting and Session Enrichment
VOLTA-AUTODOXER is a self-deleting Windows infostealer that expands standard host collection into active, token-driven session enrichment. Beyond extracting saved browser credentials, cookies, and payment cards across Chromium and Firefox engines, it targets profiles for Thunderbird, Outlook, Slack, Steam, Battle.net, Discord, Telegram, Roblox, and cloud infrastructure configuration files including `gcloud`, `OpenVPN`, and `Hide.me`.
The binary executes inline cracking against encrypted cryptocurrency vault files and dumps plaintext Wi-Fi credentials via direct calls to `WlanGetProfileList`. Following initial collection, VOLTA reuses stolen authenticated tokens directly from the victim host. It queries the victim's Gmail API for vendor receipts—specifically targeting Amazon, DoorDash, Uber Eats, Grubhub, Walmart, and PayPal—to extract verified home addresses. Stolen Uber Eats and DoorDash session tokens are queried for order history, Discord tokens are used for profile enrichment, and Roblox account values are computed via live API lookups against `Rolimons.com`.
Multi-Model Triage Architecture
The malware serializes its scraped telemetry into a local file named `Identity_Report.txt`, builds a raw JSON payload without external SDKs, and issues an HTTP POST directly to OpenRouter (`https://openrouter.ai/api/v1/chat/completions`) using a standard `Mozilla/5.0` User-Agent string over the victim's local network.
- Primary Model: `anthropic/claude-opus-4.6`
- Secondary Fallback: `moonshotai/kimi-k2`
- Tertiary Fallback: `google/gemini-2.5-flash`
The client monitors HTTP responses for refusal indicators such as "I can't" or "I'm not able" before cascading to the next provider in the chain. Request parameters enforce a hardcoded temperature of `0.2` to guarantee deterministic output. The system prompt instructs the model to act as a forensic intelligence analyst. The returned payload (`AI_Analysis.txt`) structures target value under a ranked "WHAT THEY'RE WORTH" section (low/medium/high tiers) and assesses security posture under "THEIR OPSEC", correlating personal markers like names and addresses against password patterns.
Execution Model and Exfiltration Stages
VOLTA relies entirely on third-party API endpoints and Telegram channels for operational infrastructure, maintaining no operator-owned C2 servers. The binary is distributed via a commoditized builder model containing six configurable build-time slots. Key storage relies on encryption using a SHA-256 key derived from two static constants embedded within the binary.
- `volta_log.zip`: Raw host credentials and harvested files.
- `volta_dox.zip`: Combined serialized identity telemetry (`Identity_Report.txt`) and LLM analysis (`AI_Analysis.txt`).
- `volta_doxnet.zip`: Supplementary network enrichment payloads.
Detection and Host Mitigation
Because VOLTA leverages headless execution routines to bypass Chrome's App-Bound Encryption (ABE), enforcing application control policies that restrict unauthorized binaries from spawning `chrome.exe` with the `--headless` flag neutralizes the ABE decryption vector. EDR platforms should alert on any non-developer process initiating headless browser switches.
In post-compromise scenarios where the primary binary has self-deleted, host forensics should prioritize preserving `%APPDATA%\volta_debug.txt`. This log contains timestamped records of harvested artifacts, network enrichment state, and LLM API execution success, providing audit trails required to scope credential and token revocation.