Cisco Talos uncovers Volta-Autodoxer infostealer malware with LLM-based victim analysis
Cisco Talos has described the Volta-Autodoxer malware, which, after stealing passwords, OAuth tokens, and other data, sends a profile of the victim for analysis to LLM models via OpenRouter – primarily the Claude Opus 4.6 model, with fallbacks to Kimi K2 and Gemini 2.5 Flash – in order to automatically select the most valuable targets.
Cisco Talos researchers described an infostealer malware with the internal designation "Volta," which the security community calls Volta-Autodoxer. The malware was identified using a new tool called CAIRN, which Talos developed specifically to detect hacking tools with built-in AI capabilities. According to Cisco Talos, it has not yet been confirmed that the malware has been deployed in real-world attacks. Volta operates in two steps: first, it collects stored passwords, Microsoft OAuth tokens, cookies, payment cards, crypto wallet seed phrases, and active sessions for Discord, Telegram, Steam, and Roblox. In the second round, it digs deeper – through linked accounts, such as Gmail, it looks up location, habits, and order history from services like UberEats or DoorDash, building a more detailed profile of the victim, particularly younger users with a rich digital life.
The malware sends the collected data for processing to language models via the OpenRouter service, which allows routing queries to various AI providers. According to the researchers, the default choice is the Claude Opus 4.6 model from Anthropic; in case the model's safety mechanisms reject the request, the malware has a built-in fallback in the form of the Kimi K2 model from Moonshot AI, followed by the Gemini 2.5 Flash model from Google. The malware's authors instruct the model in a 240-word prompt to act as a "forensic intelligence analyst" and produce a summary of the victim, their estimated value, and opsec level, including connecting correlations between individual pieces of data.
According to Cisco Talos researcher Ryan Fetterman, the prompt deliberately contains nothing explicitly malicious, in order to reduce the risk of the model rejecting it. AI analysis is an optional feature – the malware also works without calling the language model. Talos believes Volta may have been developed for sale to a wider circle of attackers, rather than for a single campaign.
Why it matters
Automated AI prioritization of victims speeds up attackers' decision-making, allowing them to identify the most valuable targets faster, before the victim manages to change passwords or invalidate tokens. The case, according to Cisco Talos, also shows that similar capabilities are becoming accessible even to less experienced attackers, since a commercial language model now handles the demanding analytical work for them.
Two audiences, two different impacts
What this means
For individuals
If you use the same or similar passwords across multiple services, linked OAuth logins, and download unofficial software, this type of infostealer can quickly piece together a detailed profile for the attacker from your accounts and order history.
For a business
Faster AI prioritization of stolen credentials shortens the time companies have to react after a breach – employee OAuth tokens and passwords compromised on personal devices can be exploited before they are revoked.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.