A flaw was found in vLLM. A remote client sharing the inference engine can trigger a Denial of Service (DoS) by submitting requests with prompt log probabilities enabled against a cached prefix. Because uninitialized memory is allocated for log probability tensors, cached token positions remain unpopulated, causing the tokenizer to encounter invalid values during decoding. This causes an overflow error that crashes the core engine process and terminates all concurrent tenant requests.
Exploitability
AV:NAC:LPR:LUI:NScope
S:UImpact
C:NI:NA:H6.5/CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H