Cisco Talos reported on October 31, 2026, that a Windows infostealer called VOLTA does something most stealers leave to a human: after emptying a victim's computer of passwords, sessions and wallet data, it sends a structured bundle of the stolen information to an AI model and asks the model to assess what the victim is worth. The malware calls itself VOLTA-AUTODOXER. Talos found it through CAIRN, its open-source framework for hunting AI-integrated malware, released in September 2026.
What an infostealer does
An infostealer quietly harvests credentials and sensitive data from an infected machine: saved browser passwords, cookies that keep a user logged in, payment cards, and cryptocurrency wallet information. Stolen session tokens are often more valuable than passwords, because they let an attacker step into an already-authenticated session without defeating multi-factor authentication.
VOLTA casts a wide net. Talos says it takes passwords and cookies from Chromium-based browsers, data from Outlook profiles and Slack workspaces, payment cards, crypto wallet seed phrases, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud-provider keys, Wi-Fi profiles, and screenshots. It then enriches the haul with the victim's own stolen sessions, searching Gmail for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart and PayPal to corroborate details such as a home address. The resulting bundle includes geolocation, employer and employer domain, an inferred birth year, card last-four digits, Slack workspaces, and document authorship metadata.
How the AI ranking works
The bundle is sent to an AI system through OpenRouter, a gateway routing requests to many model providers, with a roughly 240-word system prompt telling the model to act as a "forensic intelligence analyst" drafting a target assessment, using only the bundle as its source of facts. The required output has two sections: "WHAT THEY'RE WORTH," ranking valuables by monetisation potential in low, medium and high tiers, and "THEIR OPSEC," assessing security posture and password hygiene. The temperature is set low, at 0.2, for consistency.
The default model is Anthropic's Claude Opus 4.6. If a model refuses — VOLTA checks for phrases such as "I can't" — it falls back to Moonshot AI's Kimi K2, then to Google's Gemini 2.5 Flash. If the buyer supplies no OpenRouter key, the AI step is skipped and the raw dossier goes out over Telegram. When it succeeds, the assessment is saved as AI_Analysis.txt and exfiltrated with the identity report in one archive. Talos has not attributed VOLTA to a specific actor or tracked its use in the wild, but believes it is sold as a builder: buyers configure a Telegram channel and API keys and receive ranked reports back.
Why AI triage changes incident response
Infostealer-as-a-service is not new; criminal marketplaces have long sold ready-made stealers, and stolen credentials have flowed into logs sold in bulk. The bottleneck was human: someone had to sift thousands of compromised machines to find the few worth exploiting further — the finance employee, the engineer with cloud keys, the executive with valuable accounts. VOLTA automates that bottleneck. A buyer can prioritise instantly and move to account takeover, business email compromise, or ransomware staging against the highest-value targets first. A session that once sat in a log for weeks may now be acted on within hours.
The AI step also runs over the victim's own internet connection through a legitimate API gateway, not attacker-owned infrastructure, making the traffic harder to distinguish from ordinary AI use. The defensive signal may lie in unusual patterns, such as a workstation suddenly sending structured personal data to an LLM API.
What founders should take from this
Startups are disproportionately exposed: they combine valuable assets — cloud accounts, code repositories, investor and customer data — with small security teams and broad access. One compromised founder or engineer laptop can hand over OAuth tokens for email, Slack, cloud infrastructure and source control in a single sweep.
The defences are unglamorous but effective: hardware security keys for the most sensitive accounts, short-lived and narrowly scoped cloud credentials, endpoint detection that flags credential-access behaviour rather than only known malware signatures, and an incident plan that assumes a stolen session is already being ranked. Revoke sessions and rotate credentials immediately, instead of treating a stealer infection as low-priority cleanup. AI is becoming infrastructure inside malware itself, doing analyst work that once required a human. Security budgets assuming attackers still triage by hand are budgeting for the last generation of threat.
The Hacker News, GBHackers, Cybersecurity News and WIRED coverage of the Talos findings
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



