Researchers Find Infostealer's AI Fallback Refusal Detector Misses Half of Refusals

Cisco Talos published analysis on October 31 of VOLTA, a self-deleting Windows infostealer that uses a chain of three AI models to build target assessments of victims. The malware checks model replies for two fixed refusal phrases and moves to the next model when it finds one, a design Talos interprets as anticipating guardrails. A test of that exact detector against 24 hand-written replies found it caught only six of twelve genuine refusals and wrongly flagged four compliant replies. The author attributes the misses to refusals worded differently from the quoted phrases and the false positives to harmless uses of the same wording.