BlockJack paper shows vLLM cross-tenant KV redirection fault attack
Researchers have introduced BlockJack, the first KV cache address-translation fault attack in large language model inference. By corrupting block-table entries in paged serving frameworks like vLLM, the attack redirects KV cache reads to another tenant's blocks without altering model parameters or prompts, resulting in both confidentiality and integrity violations. The study demonstrates the attack's physical feasibility using Rowhammer and evaluates assignment validation as an effective defense. Open article →