CVE-2026-53923

Home/CVE/vLLM is an inference and serving engine for large language models (LLMs). From 0.5.5 until 0.23.1rc0, integer truncation

CVE

↯ PDF report

vLLM is an inference and serving engine for large language models (LLMs). From 0.5.5 until 0.23.1rc0, integer truncation

vLLM is an inference and serving engine for large language models (LLMs). From 0.5.5 until 0.23.1rc0, integer truncation of tensor dimensions in vLLM's GGUF dequantize kernels (csrc/quantization/gguf/gguf_kernel.cu) causes partial tensor processing. The output tensor is allocated at full size via torch::empty (uninitialized memory), but the dequantize CUDA kernel processes only a truncated number of elements.

The unfilled portion of the output tensor retains whatever was previously in GPU memory. In multi-tenant inference deployments, this residual GPU memory may contain tensor data from other users' inference requests, constituting information disclosure. This vulnerability is fixed in 0.23.1rc0.

Monitor

⚠ NVD has not scored this CVE yet - manual triage required (common for recent CVEs)

No Sigma yet — build one → YARA rules0

Look this up elsewhere - one-click external pivots

NVD CVE.org CISA Exploit-DB GitHub PoC VulnCheck KEV GreyNoise

▸ How to read a CVE - triage first, then detect and patch

This page is every public fact about CVE-2026-53923, cross-linked. Its job is to answer one question fast - does this need my attention now? - and then hand you the two things you do about it. Here is how an analyst reads it.

Triage: should I act now? Four signals, and they are not interchangeable:

CVSSseverity - how bad it is IF exploited, 0-10. A high CVSS alone is not urgency; a flaw can be a perfect 10 and never actually be attacked. EPSSprobability - a model’s estimate of the chance it is exploited in the next 30 days, 0-1. This is the “will it actually happen” signal. CISA KEVconfirmed - it is being exploited in the wild right now. The strongest signal on the page; KEV beats any score. Weaponisedavailability - public exploits / PoCs, and especially Metasploit modules rated Excellent / Great. Reliable, packaged exploit code means low-skill attackers can use it today.

How they combine: KEV, or a dependable Metasploit module, means patch now regardless of CVSS. High CVSS + low EPSS + no exploit is real but not an emergency - schedule it. Low CVSS but KEV-listed still gets patched now. The verdict above already weighed these for you; this is how it got there.

Then what - two workflows:

Detectwhen you cannot patch today, follow this CVE to the ATT&CK techniques it enables, then Build a SIEM detection (the green button) - author a rule, test it in Atomic, deploy it. That buys visibility while the patch waits. PatchAffected products / packages tell you if you are exposed; Fixed versions by distribution and Vendor advisories give the exact version that closes it.

Reading order for the panels below: verdict + badges, then Public exploits / Metasploit (is it weaponised), then ATT&CK techniques + Sigma / IDS rules (can I detect it), then Affected products / packages + Fixed versions (am I exposed, what patches it), then Threat actors / IOCs (who uses it), then Scoring & timeline / references (the evidence).

◆

ATT&CK techniques

Techniques this CVE enables. Pills with a solid outline are high confidence - named directly in ATT&CK or Nuclei, or human-curated by CTID; the rest are inferred from the weakness type using MITRE's CVE Mapping Methodology and the CWE → CAPEC chain. Broad, generic-weakness guesses are filtered out. A small N× marks a technique that N independent sources agree on.

▤ Build a SIEM detection for these techniques

▤

CAPEC attack patterns

Attack patterns this CVE enables - the bridge from weakness to ATT&CK technique.

CAPEC-CAPEC-116 · Excavation CAPEC-CAPEC-13 · Subverting Environment Variable Values CAPEC-CAPEC-169 · Footprinting CAPEC-CAPEC-22 · Exploiting Trust in Client CAPEC-CAPEC-224 · Fingerprinting CAPEC-CAPEC-285 · ICMP Echo Request Ping CAPEC-CAPEC-287 · TCP SYN Scan CAPEC-CAPEC-290 · Enumerate Mail Exchange (MX) Records CAPEC-CAPEC-291 · DNS Zone Transfers CAPEC-CAPEC-292 · Host Discovery CAPEC-CAPEC-293 · Traceroute Route Enumeration CAPEC-CAPEC-294 · ICMP Address Mask Request

⬡

Weakness Classification

CWE-200Exposure of Sensitive Information to an Unauthorized Actor

CWE-681Incorrect Conversion between Numeric Types

▤

Affected Packages

Language-ecosystem packages (from OSV) tied to this CVE, with the version that fixes it - the dependency-level detail NVD doesn’t carry.

PyPI vllm MODERATE

▤

Related CVEs

CVEs linked to this one by a shared weakness (CWE) or affected product - joins on data already in the engine, with the reason shown per row. Not a guess.

CVE-1999-0059 IRIX fam service allows an attacker to obtain a list of all files on the server. same CWE-200 HIGH CVE-1999-0524 ICMP information such as (1) netmask and (2) timestamp is allowed from arbitrary... same CWE-200 MEDIUM CVE-1999-1136 Vulnerability in Predictive on HP-UX 11.0 and earlier, and MPE/iX 5.5 and earlie... same CWE-200 CVE-1999-0453 An attacker can identify a CISCO device by sending a SYN packet to port 1999, wh... same CWE-200 CVE-2007-4988 Sign extension error in the ReadDIBImage function in ImageMagick before 6.3.5-9 ... same CWE-681 HIGH CVE-2007-4268 Integer signedness error in the Networking component in Apple Mac OS X 10.4 thro... same CWE-681 HIGH CVE-2008-1721 Integer signedness error in the zlib extension module in Python 2.5.2 and earlie... 1 same CWE-681 CVE-2008-3282 Integer overflow in the rtl_allocateMemory function in sal/rtl/source/alloc_glob... same CWE-681 HIGH

🔗

References & Sources

Source URLs (vendor pages, mailing lists, write-ups). Exploit/PoC links are in their own section above to avoid duplication.

https://github.com/vllm-project/vllm/commit/f219788f91952827132fa4fdf916427cd20d225e

https://github.com/vllm-project/vllm/pull/44971

https://github.com/vllm-project/vllm/security/advisories/GHSA-5jv2-g5wq-cmr4