Super Suffixes: Bypassing Text Generation Alignment and Guard Models Simultaneously

The rapid deployment of Large Language Models (LLMs) has created an urgent need for enhanced security and privacy measures in Machine Learning (ML). LLMs are increasingly being used to process untrusted text inputs and even generate executable code, often while having access to sensitive system controls. To address these security concerns, several companies have introduced guard models, which are smaller, specialized models designed to protect text generation models from adversarial or malicious inputs. In this work, we advance the study of adversarial inputs by introducing Super Suffixes, suffixes capable of overriding multiple alignment objectives across various models with different tokenization schemes. We demonstrate their effectiveness, along with our joint optimization technique, by successfully bypassing the protection mechanisms of Llama Prompt Guard 2 on five different text generation models for malicious text and code generation. To the best of our knowledge, this is the first work to reveal that Llama Prompt Guard 2 can be compromised through joint optimization. Additionally, by analyzing the changing similarity of a model's internal state to specific concept directions during token sequence processing, we propose an effective and lightweight method to detect Super Suffix attacks. We show that the cosine similarity between the residual stream and certain concept directions serves as a distinctive fingerprint of model intent. Our proposed countermeasure, DeltaGuard, significantly improves the detection of malicious prompts generated through Super Suffixes. It increases the non-benign classification rate to nearly 100%, making DeltaGuard a valuable addition to the guard model stack and enhancing robustness against adversarial prompt attacks.

Spill The Beans: Exploiting CPU Cache Side-Channels to Leak Tokens from Large Language Models

Side-channel attacks on shared hardware resources increasingly threaten confidentiality, especially with the rise of Large Language Models (LLMs). In this work, we introduce Spill The Beans, a novel application of cache side-channels to leak tokens generated by an LLM. By co-locating an attack process on the same hardware as the victim model, we flush and reload embedding vectors from the embedding layer, where each token corresponds to a unique embedding vector. When accessed during token generation, it results in a cache hit detectable by our attack on shared lower-level caches. A significant challenge is the massive size of LLMs, which, by nature of their compute intensive operation, quickly evicts embedding vectors from the cache. We address this by balancing the number of tokens monitored against the amount of information leaked. Monitoring more tokens increases potential vocabulary leakage but raises the chance of missing cache hits due to eviction; monitoring fewer tokens improves detection reliability but limits vocabulary coverage. Through extensive experimentation, we demonstrate the feasibility of leaking tokens from LLMs via cache side-channels. Our findings reveal a new vulnerability in LLM deployments, highlighting that even sophisticated models are susceptible to traditional side-channel attacks. We discuss the implications for privacy and security in LLM-serving infrastructures and suggest considerations for mitigating such threats. For proof of concept we consider two concrete attack scenarios: Our experiments show that an attacker can recover as much as 80%-90% of a high entropy API key with single shot monitoring. As for English text we can reach a 40% recovery rate with a single shot. We should note that the rate highly depends on the monitored token set and these rates can be improved by targeting more specialized output domains.