New Technique Medium TD-2026-0136 For CISO, AppSec, AI Eng, Red Team AI-assisted
BlockJack: Fault attack redirects vLLM KV cache reads across tenants
BlockJack is a fault attack on the paged KV cache address translation in LLM serving frameworks such as vLLM. By corrupting block-table entries, it redirects a request's KV cache reads to blocks assigned to another tenant's request, without modifying model parameters, prompts, or cached KV states. The authors tested seven models across Qwen, Llama, Gemma, and Phi families, and report that a faulted mapping exposes victim context.
Why it matters
Multi-tenant LLM serving hosts that share a KV cache can leak one tenant's context to another and let an attacker steer another tenant's answers. The integrity attack caused up to 41.1% of responses on a 248-scene medical benchmark to recommend less urgent care.
Do this
- Apply the vLLM and Sonar patches for block-table assignment validation once available in your deployed version.
- Enforce per-tenant isolation of KV cache blocks where your serving stack allows it, and avoid sharing cache across tenants.
- Monitor for unexpected changes in block-table mappings and for responses that reference content from other sessions.
- Assess Rowhammer-resistant memory and ECC controls on GPU or host memory that hosts LLM serving workloads.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“By corrupting block-table entries, BlockJack redirects a request's KV cache reads to blocks assigned to another tenant's request without modifying model parameters, prompts, or cached KV states.”
“Following responsible disclosure, vLLM and Sonar confirmed the vulnerability and proposed patches that we reviewed and validated.”
Sources
Affected / related
vLLM, BlockJack, Rowhammer, Sonar
Published 2026-11-19 11:31 UTC · updated 2026-11-20 04:35 UTC · AI-assisted summary, checked against quoted sources · How we score