VOLTA Infostealer Uses AI Models to Rank Stolen Victims by Payout

Cisco Talos disclosed that the VOLTA infostealer builds a dossier on each victim from stolen credentials and data, then routes it through large language models via OpenRouter to rank victims by monetization value, with fallbacks designed to get around refusals.

Infostealers have always cast a wide net. The new VOLTA malware adds a second step that makes the net smarter: after stealing everything it can, it hands the haul to an AI model and asks which victim is worth the most. Cisco Talos disclosed the malware, which calls itself VOLTA-AUTODOXER, in research shared with WIRED and published on the Talos blog, after flagging it through a new framework called CAIRN that is built to identify hacking tools incorporating AI capabilities.

VOLTA starts as a mature, self-deleting Windows infostealer. Its first sweep harvests the standard stealer haul: browser passwords, cookies, payment cards, cryptocurrency wallet seed phrases, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud-provider keys, Wi-Fi profiles and screenshots. That alone would make it a serious credential thief. Infostealers of this kind have fed some of the largest breaches in recent years, with WIRED noting that stolen logins gathered by infostealers were behind breaches including the Snowflake-linked incidents affecting AT&T and Ticketmaster and the theft of a terabyte of Disney data, and that the malware category has been used to steal hundreds of millions of credentials.

What sets VOLTA apart is what happens after the theft. The malware builds a structured evidence bundle about the person behind the machine: geolocation and home address, employer and employer domain, inferred birth year, card last four digits, Slack workspaces, document authorship metadata, iPhone backup indicators and a live Roblox account valuation. It then uses the stolen session access to enrich that bundle further, searching the victim’s Gmail for purchase receipts to corroborate a home address, pulling UberEats and DoorDash order history, and enriching Discord accounts through the stolen token. The result is less a credential log than a dossier.

That dossier is serialized and sent to OpenRouter, a service that routes requests to multiple model providers, under a system prompt instructing the model to act as a forensic intelligence analyst drafting a target assessment. The required output includes a section titled WHAT THEY ARE WORTH, with valuables ranked by monetization in low, medium and high tiers, and a section on THEIR OPSEC, an assessment of the victim’s security posture and password hygiene. In other words, the model is not being asked to break in. It is being asked to triage after the break-in, turning raw stolen data into a prioritized, citation-backed list of what to exploit first.

The engineering shows the operators expected to hit safety refusals, and built around them. VOLTA’s default model is Claude Opus 4.6, accessed through OpenRouter, with a fallback chain to Moonshot’s Kimi K2 and then Google’s Gemini 2.5 Flash. Each response is checked for refusal phrases such as I cannot and I am not able, and if one is detected, the request passes to the next model in the chain. The temperature is set to 0.2, a low setting that favors deterministic, consistent output, which is what a triage tool wants. Talos describes the failover as implying a direct engineering effort and an expectation of hitting model guardrails, not an accident.

That design is the disclosure’s most important point for defenders. AI integration in malware is often framed as helping attackers write code or craft phishing messages. VOLTA uses AI differently: as an intelligence analyst on behalf of the operator, after the theft has already happened. The novelty, Talos says, is that the AI layer is not supporting a traditional malware capability but working as a prioritization function. An operator with thousands of stolen credential sets no longer has to sort them by hand. The model does the sorting, and the fallback chain is designed to keep it sorting even when one model refuses.

There are limits to what the research establishes. Talos did not track whether VOLTA has been used in the wild, according to WIRED, so the disclosure describes a capability found in the malware, not a measured campaign with a known victim count. The malware is self-deleting and, Talos notes, has no attacker-owned infrastructure, which makes traditional takedown harder. The research also does not claim that every model in the chain complied in every case. The fallback exists precisely because refusals happen, and the public research does not publish a compliance rate for each model.

For users, the practical defenses are unchanged but more urgent. VOLTA’s haul depends on credentials and sessions stored on the infected machine, and on the second sweep through email, food delivery and chat accounts. Strong, unique passwords, hardware-backed passkeys where available, session hygiene and revoking unknown sessions all reduce what a stealer can collect in the first place. For model providers and routing services, the disclosure raises a harder question: how to detect a request that is not asking the model to do something harmful directly, but to rank the proceeds of a crime that has already been committed, and how to respond when the same request can simply be re-routed to another model.

VOLTA is unlikely to be the last malware to try this pattern. Once triage is automated, the economics of mass credential theft improve: steal widely, let AI sort, exploit the best targets first. Defenders who treat AI in malware as only a code-writing aid will miss the version that shows up as an analyst, quietly ranking victims by payout in the background.