Listen to this Post
How the CVE Works
The vulnerability arises from a poorly optimized regex pattern in vLLM’s pythonic_tool_parser.py. The regex `r”\[([a-zA-Z]+\w\(([a-zA-Z]+\w=.,\s)([a-zA-Z]+\w=.\s)?\),\s)([a-zA-Z]+\w\(([a-zA-Z]+\w=.,\s)([a-zA-Z]+\w=.\s)?\)\s)\]”` contains nested quantifiers and optional groups, leading to catastrophic backtracking. When processing malicious inputs like [A(A= )A(A=, )A(A=, )...], the regex engine enters exponential time complexity (O(2^N)), causing CPU exhaustion. This disrupts vLLM’s API, GPU memory retention, and overall service stability.
DailyCVE Form
Platform: vLLM
Version: Pre-fix versions
Vulnerability: ReDoS
Severity: Critical
Date: 2023-XX-XX
Prediction: Patch expected within 30 days.
What Undercode Say:
Exploit:
malicious_input = "[A(A=\t)A(A=,\t" 50 + "]" Triggers ReDoS import re regex = r"[([a-zA-Z]+\w(([a-zA-Z]+\w=.,\s)([a-zA-Z]+\w=.\s)?),\s)([a-zA-Z]+\w(([a-zA-Z]+\w=.,\s)([a-zA-Z]+\w=.\s)?)\s)]" re.match(regex, malicious_input) CPU spikes
Mitigation:
- Regex Optimization: Replace with a non-backtracking parser (e.g.,
re2):import re2 safe_regex = re2.compile(r"[([a-zA-Z]+\w([^)]))]") Linear time
2. Input Sanitization:
def sanitize_tool_input(input_str: str, max_length=100) -> bool:
if len(input_str) > max_length or "A(A=" in input_str:
raise ValueError("Suspicious input")
3. Rate Limiting:
Use Nginx to throttle requests limit_req_zone $binary_remote_addr zone=api_limit:10m rate=10r/s;
Detection:
Monitor CPU spikes (Linux)
top -b -n 1 | grep python | awk '{if ($9 > 90) print "ReDoS suspected"}'
Patch Analysis:
The fix (18454) reduced complexity to O(N²), but a full rewrite using AST parsing is recommended:
import ast def validate_tool_calls(input_str: str) -> bool: try: ast.parse(input_str) return True except SyntaxError: return False
GPU Memory Recovery:
Kill stuck processes holding GPU memory nvidia-smi --query-compute-apps=pid --format=csv | xargs kill -9
References:
Sources:
Reported By: github.com
Extra Source Hub:
Undercode
