CVE-2026-54232 Vulnerability Analysis & Exploit Details

CVE-2026-54232 Vulnerability Summary

CVE-2026-54232: vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.22.1, the vLLM Dockerfile is vulnerable to a dependency confusion attack through the flashinfer-jit-cache package. The package is installed from a custom index (flashinfer.ai/whl/) using --extra-index-url, but the package name was not registered on PyPI, and UV_INDEX_STRATEGY="unsafe-best-match" is set globally. An attacker who registers flashinfer-jit-cache on PyPI with version 0.6.11.post2 can execute arbitrary code as root during the Docker build and backdoor every resulting container image, enabling exfiltration of all user prompts, API credentials, and model data from production vLLM deployments This vulnerability is fixed in 0.22.1.

CVE-2026-54232 References

External References

• https://github.com/vllm-project/vllm/security/advisories/GHSA-jrf6-vqxq-pjv2

CWE Common Weakness Enumeration

CAPEC Common Attack Pattern Enumeration and Classification

• Leveraging/Manipulating Configuration File Search Paths CAPEC-38 This pattern of attack sees an adversary load a malicious resource into a program's standard path so that when a known command is executed then the system instead executes the malicious component. The adversary can either modify the search path a program uses, like a PATH variable or classpath, or they can manipulate resources on the path to point to their malicious components. J2EE applications and other component based applications that are built from multiple binaries can have very long list of dependencies to execute. If one of these libraries and/or references is controllable by the attacker then application controls can be circumvented by the attacker.
• Search Order Hijacking CAPEC-471 An adversary exploits a weakness in an application's specification of external libraries to exploit the functionality of the loader where the process loading the library searches first in the same directory in which the process binary resides and then in other directories. Exploitation of this preferential search order can allow an attacker to make the loading process load the adversary's rogue library rather than the legitimate library. This attack can be leveraged with many different libraries and with many different loading processes. No forensic trails are left in the system's registry or file system that an incorrect library had been loaded.

Other 5 Recently Published CVEs Vulnerabilities

• CVE-2026-10658 – A missing length validation in the Zephyr Bluetooth Host ISO receive path can be triggered by malformed HCI ISO data. In bt_iso_recv() (subsys/blue...
• CVE-2026-10651 – A malformed Bluetooth Classic SDP attribute can trigger a reachable assertion in Zephyr's SDP parser. In subsys/bluetooth/host/classic/sdp.c, bt_sd...
• CVE-2026-10645 – Zephyr's ext2 directory-entry parser does not fully validate on-disk directory entry structure before copying the entry name and advancing traversa...
• CVE-2026-54236 – vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.23.1rc0, the fix for CVE-2026-22778, which introduced a saniti...
• CVE-2026-54235 – vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.23.1rc0, ll temperature validation gates use comparison operat...