BlockJack redirects tenant KV cache reads in LLM serving
Original title: BlockJack: Cross-Tenant Redirection Fault Attacks on Paged KV Cache Address Translation in LLM Serving
Corrupting vLLM block-table entries exposed another tenant's context across seven model families; vendors proposed patches.
Researchers present BlockJack, a fault attack on paged KV-cache block tables in LLM serving, studied in vLLM. Corrupting entries redirects a request's cache reads to another tenant's blocks without changing weights, prompts, or cached states. Tests on seven models from the Qwen, Llama, Gemma, and Phi families exposed victim context. On a 248-scene medical benchmark, up to 41.1% of integrity-attack responses recommended less urgent care. After disclosure, vLLM and Sonar confirmed the flaw and proposed patches the authors validated.
Unconfirmed: The paper does not report exploitation outside the authors' Rowhammer demonstration.
- The attack produces confidentiality and integrity violations without editing cached KV states.
- Authors formally characterize when redirect targets in paged mappings are predictable.
- Physical feasibility was shown with Rowhammer, and assignment validation was evaluated as a defense.
Impact: cross-tenant KV reads leak context and steer answers
| Product | Affected |
|---|---|
| vLLM | — |
- 1.Review proposed vLLM and Sonar patches covering block-table assignment validation.
From the original article
1 min read at arxiv.orgLarge language model (LLM) serving frameworks accelerate inference by retaining previously computed states in a key-value (KV) cache. Paged serving frameworks such as vLLM organize this cache through block tables analogous to OS page tables, which map each request's logical KV blocks to physical KV blocks, forming a software-managed address translation layer.
By Qifan Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.