Cisco identifies malware using Claude to rank stolen assets

Automated appraisal of stolen data

Cisco Talos recently disclosed a new Windows infostealer named VOLTA-AUTODOXER. This malware is designed to steal a victim's login credentials and then feed that information into a commercial artificial intelligence model. The goal is to have the AI rank the victim's assets based on their potential resale value.

The malware acts as an automated stand-in for a human analyst. Instead of a person manually looking through stolen files to see what is worth keeping, the AI performs this triage. It creates a structured profile of the account holder and provides a target assessment that grades password habits and ranks valuables by resale tier.

According to research published by analyst Ryan Fetterman, the malware converts the gathered data into a JSON format. It then sends this bundle through a model-routing service with a specific prompt. This prompt instructs the AI to pose as a forensic intelligence analyst to draft the final assessment.

How the malware gathers intel

VOLTA-AUTODOXER is highly aggressive in the types of data it collects from an infected computer. It pulls browser passwords, cookies, saved payment cards, and cryptocurrency wallet seed phrases. It also targets sessions for popular platforms like Discord, Telegram, Steam, and Roblox, along with Microsoft sign-in tokens and cloud keys.

The malware goes beyond simple file copying for certain assets. For example, if it finds a cryptocurrency wallet, it runs an embedded cracker against the vault. It also scrapes every saved Wi-Fi password in plaintext and takes screenshots of the victim's desktop to build a more complete picture of their digital life.

To expand the profile, the malware uses stolen sessions to search Gmail for receipts from major retailers like Amazon, Walmart, and DoorDash. It uses these receipts to confirm home addresses and order histories. It even uses stolen Discord tokens to fill out additional details about the victim's identity and location.

Bypassing AI safety guardrails

One of the most notable features of VOLTA-AUTODOXER is its ability to route around AI refusals. The malware defaults to Anthropic’s Claude as its primary engine for the appraisal task. However, if Claude identifies the request as a violation of its safety guidelines and refuses, the malware is programmed to fail over to other models.

The tool follows a specific chain of three models to ensure it gets the desired output. If Claude fails, it tries Moonshot AI’s Kimi K2, and finally, it attempts to use Google’s Gemini 2.5 Flash. This design is a deliberate attempt to overcome the safety guardrails that many commercial AI companies have put in place.

The malware also fixes the AI's temperature at a low 0.2 to ensure steady and predictable output. If the attacker does not provide a specific model key, the malware simply skips the AI step entirely. In that case, it sends the raw stolen file directly over Telegram without any automated appraisal.

A product for cybercriminals

Cisco Talos assessed VOLTA-AUTODOXER as a builder kit intended for resale rather than a tool owned by a single attacker. The code includes six swappable slots for Telegram channels, keys, and settings. These slots are filled in when each individual buyer's copy of the malware is compiled for use.

The malware is designed to be difficult to track because it uses third-party services for its network traffic. This leaves no attacker-owned infrastructure for security researchers to seize or shut down. By using these external services, the developers can distribute the tool more easily to various criminal actors.

Cisco Talos framed this trend as 'effort displacement.' This means the AI is replacing the manual analytical work that a human operator would normally have to do by hand. This pattern follows two other AI-wired tools that were documented by the firm earlier this autumn.

Defense and next steps

While the firm analyzed one sample of the malware, they did not state whether it has been used against real victims yet. The malware is a self-deleting infostealer, which helps it avoid detection by staying on a system for a shorter period. It specifically uses a method called 'headless mode' to lift browser tokens from Chrome.

To defend against this type of threat, Cisco Talos recommends restricting which programs are allowed to start Chrome in headless mode. They also advise users to rotate any exposed keys and passwords immediately if they suspect a breach. Victims are encouraged to save their debug logs for further analysis by security professionals.

Security experts also warn users to move any cryptocurrency that is stored with a plaintext seed phrase to a more secure location. As AI-driven tools become more common in the cybercrime ecosystem, the speed at which attackers can profile and prioritize victims will likely increase. This makes proactive security measures more important than ever.