Skip to main content

CVE detail

CVE-2026-54234

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Prior to 0.24.0, a frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which is then converted to negative one when the engine selects the next live token for a request and is written back into the drafter's input ids; that out-of-vocabulary value is later consumed by the model's embedding and attention path and crashes the engine worker with a GPU device-side assertion. The same triggering request sequence is reachable through the public gRPC Generate and Abort endpoints, so a remote client that can send generation requests can crash the shared engine worker, aborting concurrent requests and causing a service-wide denial of service for other clients of the deployment until the worker is restarted. This issue is fixed in version 0.24.0.

CVSS 7.5 · HighBuzz score 18.9

Buzz score

Why this CVE is surfacing

Buzz score total 18.9

This all-time snapshot uses the same composite formula as Trending across a 30-year evidence window, rather than a current rolling window.

Buzz score components · mention 13.9 · diversity 5.0 · KEV 0.0 · OTX 0.0 · PoC 0.0
Mention score
13.9
3 evidence mentions in the snapshot
Diversity score
5.0
1 sources across 1 categories
KEV score
0.0
No KEV entry observed
OTX score
0.0
0 OTX pulses
PoC score
0.0
0 repos · best confidence N/A
Best PoC traction
0
Maximum stars on a matched PoC repo

Why it matters now

Mention timeline

Total mentions
3
within the 30d window
Peak daily
3
highest bucket

Evidence

Source links by recency

Newest mentions first
3 source links · newest first

Exploit code

Public exploit repository references

Public PoC repositories are third-party, potentially unsafe artifacts. Treat their code as untrusted and use it only on authorized systems in an isolated, least-privilege environment. cvebuzz does not execute the code or verify that an exploit works.

0 repository references · best confidence N/A · max 0 stars
No public PoC repositories have been matched yet.

Related records

Similar CVEs

6 related CVEs with shared weakness or product evidence