Rather than just detecting automated scams by IA, TokenTrap occupies them with unnecessary and costly calculations, without ever delivering them IDs or payment.
Scam systems built on language models are becoming more sophisticated: dialogue that adapts to the victim, memory of the conversation, external tools, several agents who divide up the work. Researchers see this as an exploitable weakness: each step of reasoning, each model call and each exchange between agents has an inference cost. Their defense, TokenTrap, consists of directing the alleged scammer to benign and worthless tasks, but requiring computational purposes, while carefully remembering passwords, bank details and other usable information.
Three bricks and guardrails
The device combines a library of tasks tailored to the targeted architecture, a router that identifies whether it is facing a single agent, a hierarchy of agents, or a leaderless collective, and a strategy that varies the presentation of requests to bypass the attacker's filters, under strict ceilings on tokens, turns, duration, and security. On 50-turn exchanges, token consumption is multiplied by 2,487, 1,616, and 568 times depending on the architecture, compared to a conversation without these tasks. The router recognizes the correct architecture in 85.67% of the 300 cases tested.
The authors present this approach as a complement to detection: short innocuous queries are enough to make an automated scam much less profitable.
Read the full article on the original source
Go to the source →