Cybersecurity researchers at Cisco Talos have uncovered a Windows infostealer malware named VOLTA-AUTODOXER that employs large language models to evaluate stolen data and prioritize victims based on their financial value. The malware collects typical credentials and session tokens, but it goes further by assembling what researchers describe as a comprehensive profile of the individual using the compromised machine. VOLTA then routes this information through artificial intelligence systems that produce a ranked assessment of each victim's worth to criminal operators.

The malware harvests passwords and cookies from Chromium browsers, payment card details, cryptocurrency wallet seed phrases, Microsoft OAuth tokens, SSH keys, cloud provider credentials, Wi-Fi profiles, and screenshots, according to the Cisco report. It also extracts data from Outlook, Slack workspaces, Discord, Telegram, Steam, and Roblox sessions. To enrich victim profiles, VOLTA accesses stolen Gmail sessions to search for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart, and PayPal, extracting delivery addresses to confirm location data. The resulting evidence bundle includes geolocation, home address, employer information, inferred birth year, partial card numbers, Slack workspace memberships, document authorship metadata, iPhone backup indicators, and live Roblox account valuations. Data exfiltration occurs over Telegram in three stages: immediately after the stealer runs, after the AI analysis completes, and following the Gmail enrichment step.

Researcher Ryan Fetterman explained that "VOLTA builds a structured evidence bundle about the person behind the machine," containing all the collected biographical and financial indicators. The malware serializes this bundle and submits it to OpenRouter, an LLM routing service, defaulting to Anthropic's Claude Opus 4.6 before falling back to Moonshot AI's Kimi K2 and Google's Gemini 2.5 Flash if earlier models refuse the request. A 240-word system prompt instructs the AI to act as a "forensic intelligence analyst" drafting a target assessment, with explicit instructions that the evidence bundle is the only permissible source of facts. According to Cisco, the required output includes a section titled "WHAT THEY'RE WORTH" that ranks valuables by monetization potential in low, medium, and high tiers, plus a section on "THEIR OPSEC" assessing the victim's security posture and password practices.

The disclosure reflects accelerating adversarial adoption of generative AI, as threat actors integrate the technology to automate analytical work that previously required manual review. VOLTA's architecture suggests the malware author may be distributing a builder kit to customers or compiling custom binaries on demand, since the malware uses a template with six slots—including Telegram channels and API keys—filled in at build time. Cisco noted that if VOLTA is commoditized, the AI dossier feature enables operators of any skill level to receive ranked victim assessments directly in their Telegram inbox without conducting their own analysis. The researchers also detailed a separate tool called NIK C2, a Go-based remote access framework that integrates OpenAI's API to analyze stolen data within compromised sessions, and notably checks victim machines for stored OpenAI API keys to fund the attacker's own AI inference—a technique called LLMjacking that lets criminals use a victim's account while evading detection. Cisco characterized NIK C2 as collapsing collection and exploitation into a single continuous session, effectively eliminating the offline processing delays that historically gave defenders critical detection windows. Organizations relying on AI-powered workflows now face a strategic dilemma: the same automation that scales legitimate operations also scales adversarial triage, potentially widening the gap between attack sophistication and defense readiness.