Why it matters
Sanket Badhe presents ScamAgent, an autonomous multi-turn framework that combines planning, conversational memory, deceptive framing, and text-to-speech to produce realistic scam calls. Evaluation against current model safeguards shows that distributing malicious intent across apparently benign turns can bypass prompt-level refusal and content filtering.
My takeaway: Test abuse across complete multi-turn trajectories, not isolated prompts. Track accumulated intent and goal decomposition, constrain high-risk impersonation, monitor memory-driven adaptation, and evaluate voice pipelines as part of the same agent-level control surface.