CVE intelligence and bounded remediation

CVE-2026-54234 — vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs

High CVSS 7.5

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.

Severity
High
CVSS
7.5 (3.1)
Published
2026-07-06
CISA KEV
Not currently listed
Ecosystem
software/application
Weaknesses
CWE-20, CWE-1284

Affected products

  • vllm / vllm

Matched remediation archetype

Resource exhaustion and denial of service

This catalog composition supplies bounded fallback guidance. Explicitly reviewed curated workflows load with the complete record below.

Check exposure

  • Identify attacker-influenced work factors including input size, nesting, compression, fan-out, regex cost, allocation, recursion, retries, and connection lifetime.
  • Map per-request and shared CPU, memory, disk, descriptor, thread, queue, and downstream-service limits.
  • Determine whether authentication, tenancy, quotas, and rate controls apply before expensive processing begins.

Remediate safely

  • Bound input size, nesting, expansion, work, concurrency, queue depth, retries, and execution time before resource-intensive processing.
  • Release resources on every success, error, cancellation, and timeout path and use backpressure instead of unbounded buffering.
  • Update affected components and add small deterministic tests that assert resource ceilings rather than exhausting a host.

Authoritative sources

Complete CVE record and remediation plan

The detailed catalog view below loads this exact record, its source evidence, and the full seven-phase agentic change plan.