VOLTA-AUTODOXER is a Windows infostealer that does something its predecessors did not: after stealing everything on the machine, it asks an AI to tell the operator which victims are worth pursuing. The malware harvests the standard haul — browser credentials, cryptocurrency wallets, Discord and Telegram tokens, SSH keys, Wi-Fi passwords — then builds a structured identity bundle containing geolocation, employer data, inferred birth year, card last-4 digits, and document metadata. It uses stolen session tokens to enrich that bundle further, querying the victim's own Gmail for Amazon and DoorDash receipts to confirm home addresses, pulling UberEats order history, and valuing Roblox inventory via live API calls. The enriched bundle is POSTed to OpenRouter under a system prompt instructing the model to act as "a forensic intelligence analyst drafting a target assessment." The required output includes a "WHAT THEY'RE WORTH" section ranking monetizable assets by low/medium/high tier, and a "THEIR OPSEC" section evaluating the victim's security posture and password hygiene. Temperature is hardcoded at 0.2 — the operator wants consistent, deterministic output, not creativity. The default model is Claude Opus 4.6, with a failover chain to Kimi K2 and then Gemini 2.5 Flash. The failover architecture is the most revealing design choice. VOLTA checks each model response for refusal phrases like "I can't" or "I'm not able" before cascading to the next model in the chain. This is not speculative engineering — it implies direct experience hitting guardrails, and a deliberate investment in refusal resilience. The malware can also operate without the LLM stage entirely if no API key is present, meaning the AI layer is an enhancement, not a dependency. The entire inference round-trip happens on the victim's own machine, over their own network connection, before exfiltration. There is no attacker-owned infrastructure. VOLTA uses a builder model with six swappable template slots filled at compile time — Telegram channel, API keys, settings — suggesting it is sold as a commodity tool. The AI dossier capability is not something the author built for personal use; it is a feature shipped to customers of any sophistication level. The stealer's enrichment pass is where the real operational innovation lives. Using the victim's own authenticated Gmail tokens, it searches for purchase receipts to corroborate physical addresses. Stolen Uber Eats and DoorDash sessions yield order histories. Discord tokens pull full account profiles. The malware doesn't just copy cryptocurrency wallets — it runs an embedded vault cracker inline and reports results. Wi-Fi passwords are dumped via direct WlanGetProfileList calls. The output is a complete identity file that an LLM can reason over. The axis of improvement here is not speed, scale, or novel attack capability. It is effort displacement. The same logic visible in the CLOSEDQUORUM implant, applied to the victim-prioritization phase rather than the attack phase. The operator no longer needs to manually sift through thousands of credential logs to identify high-value targets. The model does the analytical work, turning a wide-net approach into targeted exploitation. The operational floor drops — any buyer of the builder can receive ranked target assessments in their Telegram inbox. Defensively, much of VOLTA's token-stealing relies on bypassing Chrome's App-Bound Encryption via headless browser launches. Application control policies restricting which processes can launch Chrome with --headless would block the ABE decryption path. The malware is self-deleting, but leaves a debug log at %APPDATA%\voltadebug.txt containing a timestamped record of what was collected and whether the LLM call succeeded. In the event of compromise, preserving that file is the first forensic priority.
