VOLTA-AUTODOXER routes structured victim bundles to OpenRouter with three-model failover chain

The malware harvests browser credentials, Discord sessions, cryptocurrency seeds, and document metadata before constructing a JSON bundle. It then POSTs this payload under a forensic analyst system prompt that demands ranked monetization tiers and OPSEC assessments. Self-deletion occurs immediately after exfiltration with no command-and-control servers observed.

Talos documented explicit refusal-string checks for phrases such as "I can't" before cascading to smaller models. This architecture directly encodes expectation of guardrail triggers. The low temperature setting and hardcoded User-Agent indicate deliberate engineering for deterministic output rather than exploratory prompting.

Wider deployment of wide-net stealers paired with post-exfiltration LLM triage reduces operator workload from raw log review to selective monetization. Model providers logging prompt patterns containing "WHAT THEY'RE WORTH" sections could enable early detection clusters. The approach shifts attacker economics from targeted initial access to bulk harvesting followed by automated prioritization.

Subsequent variants are likely to incorporate additional permissive models and refine citation requirements in system prompts to improve dossier reliability.