A piece of password-stealing malware that calls itself Volta-Autodoxer finishes its job by asking a chatbot to size up the victim. Researchers have not tracked it in real attacks, but the program works, and it shows where this kind of malware is going.
The model does the reading
The malware first collects the usual things: saved passwords, session cookies, card details and the codes that keep you signed in to apps like Discord, Telegram and Microsoft accounts. Then it uses those stolen sessions to look further. It searches the victim's Gmail for receipts from shops and delivery apps, which confirm where they live, and it pulls their food-order history.
All of that goes to OpenRouter, a service that forwards a request to many different AI providers. The instructions tell the model to act as an intelligence analyst, to say what the person does and what they are worth, to rank their valuables as low, medium or high, and to judge their password habits. They also ask it to connect dots, such as checking whether a Wi-Fi password is the person's last name.
The authors expected refusals. The request goes first to Claude Opus 4.6, and if the reply sounds like a refusal, to Kimi K2, and then to Gemini 2.5 Flash. The prompt is worded to sound like ordinary analyst work. A model's guardrails slow this down, but whoever builds the malware plans for that, so the protection has to come from the accounts being stolen, not from the chatbot.
What one infection looks like now
Picture a bookkeeper who installs a cracked video editor on the laptop she also uses for client work. In the old version, the criminal gets a folder with hundreds of logins and cookies and has to decide what matters. In the new one, the first thing they read is a short report: she works at a small accounting firm, she is signed in to its cloud accounting tool and its Slack, a delivery receipt confirms her home address, and her Wi-Fi password is her surname plus a year. The firm's accounting login is marked high, her card medium, her game account low.
Reading a file like that carefully takes a person perhaps ten to twenty minutes. If a thousand machines are infected, that is about six weeks of full-time reading. The model does it for what is likely a few cents a victim, so a thousand reports cost tens of dollars.
Speed is what the attacker was missing
Stolen sessions expire and people change passwords, so a login loses value the longer it waits. The 2024 break-ins show the other side: close to 80% of the accounts the attackers used had been exposed before. Those attackers worked through old piles slowly. A tool that reads every new infection at once gets to the fresh logins while they still work, and can also go back through old piles in full.
Volta also looks built to be sold. Its settings are filled in when each copy is compiled, so a buyer enters their own details and receives finished victim reports in their Telegram messages. That means a criminal with no analyst skills gets the same ranked list as a skilled one. In September, Talos described a similar pattern: malware that hands its next moves to a chatbot. In both cases the criminal replaces the step that needed the most sustained human attention with a model call. The next step to go will probably be the ransom: the report already says what the victim is worth, so the demand can be set from it.
The practical answer is dull and works. A second sign-in step on your work and email accounts means an old stolen password alone no longer opens them. Keeping work accounts off any computer used for games or cracked software removes the file that gets ranked highest. If you ever suspect an infection, signing out of all devices matters as much as changing the password, because it is the live sessions that this malware takes.