Attacks on AI Research AI security importance 45 1 min read In your stack
From Token Changes to Task Failures: Bit Flips in a Shared KV-Cache Block
Fault injection into vLLM serving Qwen3-8B shows that a single bit flip in a shared KV-cache block often produces wrong answers across tasks.
Paper
Sources
arXiv (AI security papers) primary
Affects and mentions
Add any of these to your stack to follow them.
Related stories
- BlockJack: Cross-Tenant Redirection Fault Attacks on Paged KV Cache Address Translation in LLM Serving arXiv (AI security papers)
- CVE-2026-130686: vLLM before 0.33.0 lacks request-wide media budget, enabling resource exhaustion NVD (AI keyword search)
- CVE-2026-130685: vLLM 0.28.0 to before 0.33.0 unauthenticated EngineCore crash via out-of-range token IDs NVD (AI keyword search)
- CVE-2026-130684: vLLM before 0.33.0 path traversal via speculators_config verifier.name_or_path NVD (AI keyword search)
One-line summary and action are AI-assisted. Read the source before acting.