New AI Security Vulnerability ‘Mistletoe’ Exposed in Speculative Decoding Techniques
Researchers have identified a novel security vulnerability in speculative decoding techniques used to accelerate large language model (LLM) inference, according to a preprint study published on arXiv. The attack, dubbed "Mistletoe," exploits weaknesses in how speculative decoding systems verify candidate tokens, potentially enabling attackers to disrupt AI performance by reducing draft-token acceptance rates.
Speculative decoding, widely adopted by US-based companies including OpenAI, Google, and Meta, allows LLMs to generate multiple candidate tokens in parallel before verification. The study demonstrates that attackers can manipulate this process through "acceleration-collapse" techniques, significantly degrading system efficiency. "The vulnerability lies in the mechanism itself," the paper states, noting that the average accepted length ($\tau$) becomes a critical attack vector.
The research team, which remains anonymous in the arXiv submission, emphasizes that the attack works across different speculative decoding implementations, including speculative sampling and parallel sampling. While the paper does not disclose specific exploit scenarios, it warns that the vulnerability could be weaponized to create denial-of-service conditions or bias output through token manipulation.
