By weakness (CWE)

CWE-681: related vulnerabilities

CVEs classified under CWE-681. Understanding the weakness class helps prioritize systemic fixes over one-off patches.

1 published vulnerability

• CVE-2026-53923HIGH 7.5

  vLLM, a popular inference engine for large language models, contains a memory leak vulnerability in how it processes quantized model weights. When the system allocates GPU memory for output tensors, a mathematical error causes the processing kernel to only fill part of the allocated space, leaving the remainder untouched. In shared multi-user deployments, this unfilled memory may contain sensitive inference data from other users' requests—potentially exposing prompts, model outputs, or intermediate computations. The flaw affects all versions from 0.5.5 through 0.23.0, and is resolved in version 0.23.1rc0 and later.