Coercing LLMs to do and reveal (almost) anything
This work highlights security vulnerabilities in LLMs that could affect developers and users, though it appears incremental by expanding on known jailbreaking issues.
The paper investigates a broad spectrum of adversarial attacks on large language models (LLMs) beyond jailbreaking, including misdirection, control, denial-of-service, and data extraction, and finds that many stem from pre-training with coding capabilities and glitch tokens.
It has recently been shown that adversarial attacks on large language models (LLMs) can "jailbreak" the model into making harmful statements. In this work, we argue that the spectrum of adversarial attacks on LLMs is much larger than merely jailbreaking. We provide a broad overview of possible attack surfaces and attack goals. Based on a series of concrete examples, we discuss, categorize and systematize attacks that coerce varied unintended behaviors, such as misdirection, model control, denial-of-service, or data extraction. We analyze these attacks in controlled experiments, and find that many of them stem from the practice of pre-training LLMs with coding capabilities, as well as the continued existence of strange "glitch" tokens in common LLM vocabularies that should be removed for security reasons.