Monday

This recording of the Ground Truth track's Monday block strings together several separate talks rather than one presentation. It opens with voice phishing: a large survey experiment compares AI voice models with human scam callers. The findings suggest that message content matters more than how human a voice sounds, and that some automated calls already look profitable, though they cost more per minute than email.

The middle sessions cover trust-based scams, where a system uses AI personas to engage inbound scammers and map conversation stages before a payment request. Another talk examines exploitation timing and the EPSS scoring model, arguing that prioritizing by probability works better than severity scores. A further session presents a BERT-based classifier that separates harmful goals from the framing used to disguise them.

The final talks address a vulnerability pipeline that silently drops findings when joins fail, and a study of AI agents attempting CTF challenges. The first shows how left joins, fallback matching and evidence packets keep unowned assets visible. The second finds that agents often succeed through pattern matching, and that open-weight models trail closed ones by less than expected at far lower cost.