TokenTrap drives up token use in LLM-based scam systems
Original title: Your Token Is Gone: A Token-Consumption Countermeasure for Disrupting LLM-Based Scam Systems
Over 50-round traces, consumption rose as much as 2,487 times versus a no-token-task baseline.
Researchers present TokenTrap, a defense that steers suspected LLM scam systems into benign, low-value computation while withholding credentials and payments. It pairs an architecture-specific token-task library with a workflow router and bounded limits on tokens, rounds, time, and safety. Against single-agent, hierarchical, and flat multi-agent workflows, reported amplification was 2,487.0×, 1,616.0×, and 567.8×. The router reached 85.67% accuracy on 300 architecture-labeled prompts.
- TokenTrap varies task presentation under explicit token, round, time, and safety limits.
- The router chooses among single-agent, hierarchical, and flat multi-agent workflows.
- Short non-sensitive requests are intended to raise context and coordination cost.
- The paper frames the method as a complement to scam detection.
From the original article
1 min read at arxiv.orgLLM-based scam systems increasingly combine adaptive dialogue, persistent context, external tools, and multi-agent coordination to automate personalized fraud. These capabilities also create a resource dependency: every reasoning step, model call, and coordination round incurs inference cost.
By Fudu Xing · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.