AI-Driven Malware Uses Claude to Rank Victims by Value Before Self-Deleting

Cisco Talos uncovered VOLTA-AUTODOXER, a Windows infostealer that steals credentials, cryptocurrency wallets, and personal data before feeding it to AI models like Claude to generate target assessments. The self-deleting malware ranks victims by monetizable assets and exfiltrates dossiers via Telegram, signaling a shift toward AI-powered cybercrime operations.

AI-Driven Malware Transforms Stolen Data Into Ranked Target Assessments

Cybersecurity researchers at Cisco Talos discovered VOLTA-AUTODOXER, an infostealer malware that uses large language models to analyze stolen data and rank victims by their financial value

1

2

. This Windows malware represents a concerning evolution in cybercrime operations, where adversarial use of generative AI shifts from automation to strategic victim profiling. VOLTA-AUTODOXER steals browser passwords, cookies, payment cards, cryptocurrency wallets, Microsoft OAuth tokens, SSH keys, cloud-provider configuration files, and screenshots before building a comprehensive dossier on each victim

1

3

.

Source: Tom's Hardware

Source: Tom's Hardware

How VOLTA Builds Structured Evidence Bundles on Victims

The infostealer malware goes beyond standard credential theft by conducting live enrichment with stolen sessions. VOLTA searches victims' Gmail accounts for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart, and PayPal to confirm home addresses

1

2

. The malware pulls order histories, reuses stolen Discord tokens to fill in profile information, and values Roblox items through Rolimons. VOLTA assembles geolocation data, employer domains, inferred birth years, the last four digits of payment cards, Slack workspaces, document authorship metadata, and iPhone backup indicators into a structured evidence bundle

2

3

. This forensic intelligence approach transforms passive stolen data into actionable profiles for attackers.

Multi-Model Architecture Ensures AI Analysis Success

VOLTA-AUTODOXER serializes the evidence bundle to JSON and sends it to OpenRouter, a unified API gateway that provides access to hundreds of AI models

1

3

. The malware defaults to Anthropic's Claude Opus 4.6 with a temperature setting of 0.2, prioritizing predictable output over creative responses

1

. When Claude refuses to comply, VOLTA switches to Moonshot AI's Kimi K2, then Google's Gemini 2.5 Flash

2

3

. The malware checks responses for refusal phrases like "I can't" or "I'm not able" before falling back to the next model. If no OpenRouter API key is provided, the self-deleting malware can still exfiltrate raw data via Telegram without AI analysis

1

.

Source: TechRadar

Source: TechRadar

AI Generates Target Assessments with Monetizable Assets Rankings

The 240-word system prompt instructs the AI to act as a "forensic intelligence analyst" drafting target assessments based solely on the evidence bundle

2

. According to Cisco Talos researcher Ryan Fetterman, the prompt demands output that is "blunt, specific, accurate" and includes a section titled "WHAT THEY'RE WORTH" with valuables ranked by monetization potential using low, medium, and high tiers

1

2

. Another required section, "THEIR OPSEC," assesses the victim's security posture and password hygiene. The AI response is saved as AI_Analysis.txt alongside Identity_Report.txt and exfiltrated in a volta_dox.zip archive file

1

. Data exfiltration occurs at three stages via Telegram: after initial theft, after AI analysis, and after live enrichment with Gmail receipts and order histories

2

.

Source: Hacker News

Source: Hacker News

Evidence Points to Commoditized Malware-as-a-Service Model

VOLTA-AUTODOXER appears designed for resale with a template containing six swappable slots filled at build time

1

3

. Customers supply custom Telegram channels, API keys, and settings, receiving a compiled binary tailored to their specifications. Cisco Talos notes this builder model enables customers of any sophistication level to receive ranked target assessments in their Telegram inbox without relying on attacker-owned infrastructure

1

2

. The malware bypasses Chrome's App-Bound Encryption by launching Chrome headless without a GUI, though application-control policies limiting headless Chrome processes can block this attack vector

1

.

Implications for Ransomware Attacks and Future Cybercrime Operations

Zscaler ThreatLabz warns that generative AI technologies excel at examining massive amounts of data for classification and content searching, making them ideal for ransomware attacks where actors need to identify the most sensitive data for theft or encryption

2

. VOLTA represents a shift where AI relieves operators of analytical work rather than improving speed or scale. Cisco Talos emphasizes that VOLTA's AI use is narrow compared to what's possible with fully AI-assisted malware, especially as models could be utilized for additional attack phases

1

. Refusal resilience may become a standard design requirement as AI integration in malware matures. After infection, victims should revoke all API keys and tokens, change passwords, and move cryptocurrency funds to new wallets with fresh seeds generated on air-gapped devices

1

3

. A volta_debug.txt file typically remains on infected systems, showing what was collected and whether the AI analysis succeeded

1

3

.

© 2026 TheOutpost.AI All rights reserved