Blog
Notes from the Faultline team on serving-layer failures in LLM inference systems.
2026-06-17
How GRIEF works
Most fuzzers for ML systems mutate prompts. GRIEF mutates timed multi-request traces — the abstraction that lets it reach KV-cache isolation failures, throughput-collapse pathologies, and scheduler invariant violations that no single prompt can trigger. A walkthrough of the architecture and the three-stage confirmation pipeline.
grieffuzzingllm-servingtrace-mutation2026-06-17
A KV-cache that forgets to forget
vLLM's V1 engine reuses freed KV-cache blocks without zeroing them — but only for non-Mamba models. A one-line gate silently breaks tenant isolation across co-scheduled requests, producing fluent, confidently wrong outputs at temperature=0. CVE-2026-7141.
vllmkv-cacheisolationcve-2026-7141