EN PT-BR
← Voltar ao feed
medium

VOLTA-AUTODOXER infostealer uses LLMs via OpenRouter to rank stolen victim data by monetization value

1 fonte(s) · 1 atualização(ões) · publicado 02/11/2026 11:47 · coletado há 1 hora

Cisco Talos disclosed a new Windows infostealer, VOLTA-AUTODOXER (VOLTA), found through its CAIRN research effort tracking AI-enabled malware. After a conventional theft phase, it sends the collected data to an LLM through the OpenRouter API. The model is asked to act as a forensic analyst, rank the victim's assets by how easily they can be monetized, and assess the victim's password hygiene. The malware deletes itself after running and uses only third-party endpoints, so there is no attacker-owned infrastructure to take down. Talos sees the same pattern, replacing the most human-intensive step with a model call, in the CLOSEDQUORUM implant.

Detalhes técnicos

  • Collection: browser passwords, cookies and saved cards; cryptocurrency wallet seed phrases; Microsoft OAuth tokens; SSH and cloud provider keys; screenshots; Discord, Telegram, Steam and Roblox sessions; plaintext passwords of all saved Wi-Fi networks. It also includes a built-in cracker for cryptocurrency wallet vaults.
  • Enrichment: builds an evidence bundle with geolocation and home address, employer, inferred birth year, card last-4 digits, Slack workspaces, document metadata, iPhone backup indicators and a Roblox account valuation. Stolen sessions are used to search Gmail for purchase receipts, pull UberEats and DoorDash order history, and enrich Discord accounts via the stolen token.
  • AI analysis: data is assembled locally in Identity_Report.txt, serialized as JSON and POSTed to OpenRouter with a system prompt. The reply is saved as AI_Analysis.txt, and both files are exfiltrated in volta_dox.zip. Assets are tagged low, medium or high, and the model looks for links between personal details and chosen passwords.
  • Model fallback: the default model is Anthropic Claude Opus 4.6. Replies are scanned for refusal wording, and on refusal the job falls to Moonshot AI Kimi K2 and then Google Gemini 2.5 Flash. Talos reads this as the operator expecting refusals and falling back to more permissive models.
  • Builder model: configuration (Telegram channel, API keys) sits in six template slots filled at build time, with weakly encrypted credentials. This suggests the author sells a builder and compiles a custom binary per customer, delivering target assessments to the buyer's Telegram inbox.

Remediação

The article gives no specific mitigation. Defenders should expect AI-assisted triage of stolen data to spread to other phases of attacks.

Linha do tempo

  1. helpnetsecurity.com ↗ · pub. 02/11/2026 11:47 · coletado 02/11 12:01 primeiro registro

    VOLTA-AUTODOXER infostealer uses LLMs via OpenRouter to rank stolen victim data by monetization value

    ler a análise desta fonte

    Cisco Talos disclosed a new Windows infostealer, VOLTA-AUTODOXER (VOLTA), found through its CAIRN research effort tracking AI-enabled malware. After a conventional theft phase, it sends the collected data to an LLM through the OpenRouter API. The model is asked to act as a forensic analyst, rank the victim's assets by how easily they can be monetized, and assess the victim's password hygiene. The malware deletes itself after running and uses only third-party endpoints, so there is no attacker-owned infrastructure to take down. Talos sees the same pattern, replacing the most human-intensive step with a model call, in the CLOSEDQUORUM implant.

    Technical Details

    • Collection: browser passwords, cookies and saved cards; cryptocurrency wallet seed phrases; Microsoft OAuth tokens; SSH and cloud provider keys; screenshots; Discord, Telegram, Steam and Roblox sessions; plaintext passwords of all saved Wi-Fi networks. It also includes a built-in cracker for cryptocurrency wallet vaults.
    • Enrichment: builds an evidence bundle with geolocation and home address, employer, inferred birth year, card last-4 digits, Slack workspaces, document metadata, iPhone backup indicators and a Roblox account valuation. Stolen sessions are used to search Gmail for purchase receipts, pull UberEats and DoorDash order history, and enrich Discord accounts via the stolen token.
    • AI analysis: data is assembled locally in Identity_Report.txt, serialized as JSON and POSTed to OpenRouter with a system prompt. The reply is saved as AI_Analysis.txt, and both files are exfiltrated in volta_dox.zip. Assets are tagged low, medium or high, and the model looks for links between personal details and chosen passwords.
    • Model fallback: the default model is Anthropic Claude Opus 4.6. Replies are scanned for refusal wording, and on refusal the job falls to Moonshot AI Kimi K2 and then Google Gemini 2.5 Flash. Talos reads this as the operator expecting refusals and falling back to more permissive models.
    • Builder model: configuration (Telegram channel, API keys) sits in six template slots filled at build time, with weakly encrypted credentials. This suggests the author sells a builder and compiles a custom binary per customer, delivering target assessments to the buyer's Telegram inbox.

    Remediation

    The article gives no specific mitigation. Defenders should expect AI-assisted triage of stolen data to spread to other phases of attacks.