Large language models are notoriously expensive to run, and they can be tricked into ignoring safety rules by jailbreak prompts. A new preprint on arXiv examines a possible link between these two problems, focusing on depth-based dynamic inference—a technique that adjusts how many layers of the network are used for a given input.
The authors position the work as a response to LLM inefficiency, but they also flag jailbreak vulnerability as a key concern. Because safety alignment is typically learned across the full network, reducing depth could create new attack surfaces. The paper explores whether such efficiency measures inadvertently make models more prone to producing harmful responses.
The released abstract is brief and stops short of describing the proposed approach or any results. As a preprint, it serves as an early signal that efficiency and safety need to be studied together rather than in isolation.