GHSA-j828-28rj-hfhp: 
vLLM vulnerability analysis and mitigation

A series of Regular Expression Denial of Service (ReDoS) vulnerabilities were identified in vLLM versions >= 0.6.3 and < 0.9.0. The vulnerability, discovered in May 2025, affects the vllm package distributed via pip and involves multiple regular expressions in the codebase that are susceptible to catastrophic backtracking when processing maliciously crafted input (GitHub Advisory).

The vulnerability affects multiple files in the vLLM codebase with four specific instances identified: 1) In lora/utils.py, the regex pattern r"((.?))$?$" for matching parenthesized content, 2) In phi4minitoolparser.py, the pattern r'functools[(.?)]' with re.DOTALL flag, 3) In servingchat.py, the pattern r'."parameters":\s(.)' for parameter matching, and 4) In benchmarkservingstructuredoutput.py, the pattern r'{.*}' for JSON extraction. The vulnerability has a CVSS v3.1 score of 4.3 (Moderate), with base metrics indicating Network attack vector, Low attack complexity, Low privileges required, and No user interaction needed (GitHub Advisory).

When exploited, these vulnerabilities can lead to severe performance degradation of the vLLM service through catastrophic backtracking in regular expression processing. This primarily affects the availability of the service, particularly when processing maliciously crafted input, though there are no direct impacts on confidentiality or integrity (GitHub Advisory).

The vulnerability has been patched in vLLM version 0.9.0 by migrating to the REGEX library instead of the standard re library. For affected versions, recommended mitigations include: limiting input string lengths, using non-recursive matching approaches, implementing stricter content constraints, and preferring structured parsing over regex for complex text processing. For JSON parsing specifically, it's recommended to use proper JSON parsing tools rather than regex (GitHub Advisory).