A Black-Box Audit of Provider-Side Token Inflation in LLM Services (arxiv.org)

A recent study introduces a concerning phenomenon known as Provider-Side Token Inflation Attack (PTIA) in pay-per-token Large Language Model (LLM) services. This manipulation technique allows dishonest providers to covertly inflate output tokens, increasing costs for users while maintaining task utility. The research outlines five distinct attack methods within the provider-controlled generation process, revealing that these methods can extend output length by over 10 times the normal baseline. This excessive financial burden poses a significant ethical dilemma for the AI/ML community, potentially undermining trust in LLM services. To combat PTIA, the researchers developed a novel lightweight audit method that requires no trusted local reference model, making detection both efficient and challenging for providers to evade. Their approach, featuring a controlled lengthening intervention, achieved an impressive average detection rate of 85.1% across four open-weight models while keeping false positives below 2%. The findings, flagging seven out of fifteen examined LLM API services for PTIA-related behavior, underscore the urgent need for transparency in LLM operations, ensuring users are not unwittingly exploited in tokenized pricing models.