VOLTA malware hands its analyst work to an LLM, and the step it replaced is the one that took human attention

Talos published an analysis on Saturday of VOLTA, a self-deleting Windows infostealer that hands its post-theft paperwork to a large language model. Researcher Ryan Fetterman sets out the sequence. The malware harvests browser passwords, cookies, payment cards, wallet seed phrases, session tokens and cloud keys. It builds an evidence bundle about the person behind the machine and posts it to OpenRouter. The system prompt tells the model to act as "a forensic intelligence analyst drafting a target assessment". The output includes a section titled "WHAT THEY'RE WORTH", which ranks the victim's assets by monetisation value in low, medium and high tiers. The default model is Claude Opus 4.6. If it refuses, VOLTA checks the reply for phrases such as "I can't" and moves down a chain to Kimi K2 and then Gemini 2.5 Flash.

The job the model took

The line worth keeping is in Talos's own conclusion. The AI layer exists "for one purpose: relieving the operator of analytical work." The researchers call this effort displacement and state the logic plainly: "Identify the step that requires the most sustained human attention and replace it with a model call."

Read that as a labour-market finding. Credential theft has been automated for years. The judgement stayed manual: which address is real, which password pattern links back to a name or a university. VOLTA hands that triage to a model at a temperature of 0.2, a setting Talos reads as the attacker prioritising accuracy and consistency. The prompt asks the model to stretch sparse, scattered records into useful inferences. That is analyst work, and it now costs one API call.

The deployment model turns it into a product. Talos found a template with six swappable slots filled at build time, which points to a builder sold to customers. In Talos's words, that makes the dossier "a feature which enables a customer of any level of sophistication" to receive ranked target assessments in a Telegram inbox. When the analysis ships inside the binary, running the operation takes much less skill.

Legitimate work follows the same pattern. Every role has a step that takes the most sustained attention. VOLTA's author decided that step was worth paying a model to do.

Your digital exhaust is the dossier

The enrichment pass is the second finding. VOLTA uses the victim's own Gmail tokens to search for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal, and uses them to confirm a home address. It pulls food-delivery order history, re-uses the Discord token and values a Roblox inventory through a live lookup. The bundle also carries the victim's employer and employer domain, their Slack workspaces and document authorship metadata. The inference round-trip with the model runs over the victim's own network connection before the result is exfiltrated. Talos found no attacker-owned infrastructure.

For a professional, that means a model can join up your employer, your workspaces and the receipts in your inbox into one file. One logged-in laptop is now worth more to whoever steals it.

Your Next Move

• Find the step in your week that takes the most sustained attention, and ask whether a model could do the first pass on your instructions. If it could, make that model call yourself and spend your time on the decision it feeds. You can sort a week's work task by task with the method for finding out whether AI will take your job.
• Audit your sessions this month. Talos's countermeasures: after any compromise, revoke tokens and change passwords. Keep credentials in hardware-backed stores. Treat any wallet whose seed phrase sat in plaintext on disk as fully compromised. If you run IT, restrict which processes can launch Chrome with --headless. Talos says this blocks the App-Bound Encryption bypass VOLTA relies on.
• Keep work and personal sessions in separate browser profiles or on separate devices. Then one stolen profile holds less of your employer, your address and your purchases in one place.