Modern large language model (LLM) systems often reuse key-value (KV) caches to speed up inference. According to a new arXiv paper, recent work has also shown that reusing KV caches from reasoning traces can improve task performance, in a manner similar to retrieval-augmented generation. This practice, however, may come with a hidden cost.
The paper, titled "Stealthy Poisoning Attacks via Reused KV Cache of Reasoning," highlights a security vulnerability in this reuse mechanism. The abstract indicates that an attacker could potentially poison the cached reasoning traces, leading to stealthy manipulation of model outputs. The exact method is not detailed in the available abstract, but the authors frame it as a new attack vector that exploits the trust placed in reused caches.
The findings underscore a trade-off between efficiency and security in LLM systems. As cache reuse becomes more common for performance gains, the paper suggests that defenses must evolve to protect against poisoning of these shared components. The abstract is truncated, so further details on the attack's implementation and mitigation are not yet available from this source alone.