ai_security▌

← feed

NOTABLE

BlockJack fault attacks on KV-cache block tables cross tenant boundaries in paged LLM serving

First fault-attack work on the address-translation layer of paged LLM serving (vLLM-style block tables): corrupting block-table entries redirects one request's KV-cache reads to blocks assigned to another tenant without modifying model weights, prompts, or cached state. The paper demonstrates both confidentiality violation (victim content appearing in attacker responses) and integrity violation (attacker-supplied context steering victim answers) on paged serving frameworks.

released 2026-11-20 · ingested 2026-11-20T12:20Z

Sources

Primary source

Cite this

APA

AI Security Index. (2026). BlockJack fault attacks on KV-cache block tables cross tenant boundaries in paged LLM serving. AI Security Index. https://arxiv.org/abs/2611.25945

BibTeX

@misc{aisec-2026-11-20-arxiv-2611-25945-blockjack-kv-cache-cross-tenant,
  title  = {BlockJack fault attacks on KV-cache block tables cross tenant boundaries in paged LLM serving},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2611.25945}},
  note   = {AI Security Index item 2026-11-20-arxiv-2611-25945-blockjack-kv-cache-cross-tenant}
}