vLLM Security Advisory: ReDoS Vulnerability in lm-format-enforcer Backend

According to the GitHub Security Advisory database, vLLM has a medium-severity vulnerability (GHSA-48jh-3gj7-fg8v) in its lm-format-enforcer backend. The issue stems from the absence of timeout and validation mechanisms when compiling attacker-supplied regular expressions in the structured_outputs.regex parameter. This flaw can lead to a Regular Expression Denial of Service (ReDoS), where a single malicious request containing a catastrophic regex can hang the structured-output compile step, stalling the engine worker and causing a denial of service.

The advisory notes that similar vulnerabilities in sibling backends, xgrammar and outlines, were addressed in a previous fix (GHSA-rwxx-mrjm-wc2m) by wrapping regex compilation with `compile_regex_with_timeout` and adding validation checks. However, the lm-format-enforcer backend remains unpatched, directly invoking `interegular` FSM construction without safeguards. Reproduction tests confirm that an attacker regex like `(a{1,300}){300}` can cause the engine to hang indefinitely, impacting concurrent requests and leading to a worker-level DoS.

Operators should carefully review their structured-outputs configuration before upgrading or deploying vLLM, especially if the lm-format-enforcer backend is in use. Any service relying on this backend for structured outputs is at risk of exploitation. If feasible, consider disabling structured outputs or switching to a backend that has been patched, such as xgrammar or outlines, until a fix is released. This vulnerability highlights the importance of ensuring consistent security measures across all components of a system, particularly when dealing with user-supplied inputs.