BlockJack: Cross-Tenant Redirection Fault Attacks on Paged KV Cache Address Translation in LLM Serving
- Published
- Record updated
Summary
BlockJack is a fault attack on the paged KV cache of LLM serving frameworks such as vLLM. By corrupting block-table entries with Rowhammer, it redirects a request's KV cache reads to blocks belonging to another tenant, without modifying model weights, prompts or cached KV states. The authors report victim information leaking into attacker responses and, on a 248-scene medical benchmark, up to 41.1% of responses recommending less urgent care, and they say vLLM and Sonar confirmed the vulnerability and proposed patches that were reviewed and validated.
Mitigation
Assignment validation was evaluated as a defense. Following responsible disclosure, vLLM and Sonar confirmed the vulnerability and proposed patches, which the authors reviewed and validated. The source does not give specific fixed versions.
Topics
Related items
- MediumCVE-2026-54026: LibreChat message injection via conversationId in request bodySimilar attack · NVD/CVE Database
- MediumCVE-2026-54046: LibreChat unauthenticated account enumeration via /api/user/verifySimilar attack · NVD/CVE Database
- HighCVE-2026-54044: LibreChat credential exposure through agent action domain repointingSimilar attack · NVD/CVE Database
- MediumCVE-2026-54043: LibreChat is an enhanced ChatGPT clone that supports multiple AI providers. Prior to 0.8.7…Similar attack · NVD/CVE Database
- MediumCVE-2026-54042: LibreChat 2FA verification unlimited code guessing via temp token endpointSimilar attack · NVD/CVE Database