Researchers have discovered a novel method to extract sensitive information from large language models, specifically targeting biases introduced by floating-point arithmetic. This technique, detailed in a new arXiv paper, exploits numerical side channels to recover relative logits and even the exact logit vector of a model, which are typically withheld by API providers. The method has been successfully demonstrated on various model checkpoints, including Qwen3-0.6B, and can recover complete output-head representations. A proposed mitigation involves subtracting the maximum logit before bias addition to suppress disclosure. AI
IMPACT This research highlights a potential vulnerability in LLM APIs, suggesting that even withheld logits can be inferred, impacting model security and privacy.
RANK_REASON Research paper detailing a novel technical method for extracting information from LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
