CVE-2025-66448
[vllm] Remote code execution via transformers_utils/get_config
Executive Summary
CVE-2025-66448 is a high severity vulnerability affecting appsec, ai-code, binary-analysis. It is classified as Code Injection. Ensure your systems and dependencies are patched immediately to mitigate exposure risks.
Precogs AI Insight
"The root cause of this vulnerability lies in within ### Summary, allowing the mishandling of memory allocation boundaries. An attacker can craft a specific payload to bypass intended access controls, establishing a persistent foothold. By intercepting insecure data flows from user input directly to rendering sinks, Precogs is designed to harden the environment against lateral movement."
What is this vulnerability?
CVE-2025-66448 is categorized as a critical Code Injection / RCE flaw. Based on our vulnerability intelligence, this issue occurs when the application fails to securely handle untrusted data boundaries.
Summary
vllm has a critical remote code execution vector in a config class named Nemotron_Nano_VL_Config. When vllm loads a model config that .
This architectural defect enables adversaries to bypass intended security controls, directly manipulating the application's execution state or data layer. Immediate strategic intervention is required.
Risk Assessment
| Metric | Value |
|---|---|
| CVSS Base Score | 8 (HIGH) |
| Vector String | N/A |
| Published | December 1, 2025 |
| Last Modified | December 1, 2025 |
| Related CWEs | CWE-94 |
Impact on Systems
✅ Remote Code Execution: Attackers achieve arbitrary command execution within the context of the application server.
✅ Privilege Escalation: Initial code execution can be exploited to pivot and elevate privileges across the network.
✅ Persistent Backdoors: Attackers can bind reverse shells, modify source files, or inject persistent access mechanisms.
How to fix this issue?
Implement the following strategic mitigations immediately to eliminate the attack surface.
1. Remove Dynamic Evaluation Completely eliminate the use of dynamic evaluation functions (eval(), exec(), system()) on untrusted input.
2. Sandboxing If dynamic execution is an absolute business requirement, isolate the execution environment in tightly constrained, non-networked sandboxes (e.g., restricted WebAssembly or isolated containers).
3. Network Segmentation Restrict outbound traffic from the application server (egress filtering) to prevent reverse shell connections.
Vulnerability Signature
// Vulnerable Node.js Execution
const exec = require('child_process').exec;
const user_domain = req.query.domain;
// VULNERABLE: Injecting user input directly into system shell commands
exec('ping -c 4 ' + user_domain, (error, stdout, stderr) =\> \{
res.send(stdout);
\});
// EXPLOIT PAYLOAD: precogs.ai ; cat /etc/passwd
References and Sources
- NVD — CVE-2025-66448
- MITRE — CVE-2025-66448
- CWE-94 — MITRE CWE
- CWE-94 Details
- Application Security Vulnerabilities
- AI Code Security Vulnerabilities
- Binary Analysis Vulnerabilities
Vulnerability Code Signature
Attack Data Flow
| Stage | Detail |
|---|---|
| Source | Untrusted payload via API or file upload |
| Vector | Input passed to a dynamic code evaluation function |
| Sink | eval(), exec(), or similar unsafe execution sink |
| Impact | Remote Code Execution (RCE), full system compromise |
Vulnerable Code Pattern
# ❌ VULNERABLE: Dynamic code evaluation
def process_data(user_input):
# Taint sink: arbitrary code execution
result = eval(user_input)
return result
Secure Code Pattern
# ✅ SECURE: Safe parsing
import ast
def process_data(user_input):
# Sanitized parsing: only evaluates literal structures
result = ast.literal_eval(user_input)
return result
How Precogs Detects This
Precogs AI Analysis Engine identifies unsafe dynamic code evaluation paths by tracking untrusted data into sinks like eval() and exec().\n