CVE-2026-55574: vLLM: unbounded regex compile hangs inference worker (ReDoS)

Prior to 0.24.0, vLLM's structured_outputs.regex parameter passes attacker-controlled regular expressions straight into the grammar compiler with no compilation timeout, and neither the xgrammar nor outlines backend performs complexity analysis on nested quantifiers, so a single crafted request triggers catastrophic backtracking-style exponential state-space expansion and hangs the worker indefinitely. There's no CVSS/EPSS scoring yet and no public exploit or Nuclei template, so exploitation likelihood is currently unquantified, but the bug requires only unauthenticated API access to any endpoint that exposes structured-output/regex-constrained generation — a common pattern for JSON-mode and tool-calling deployments. The blast radius depends on how many downstream services expose this parameter to end users versus keep it server-side only; vLLM is a widely deployed inference engine so any multi-tenant or public-facing serving stack accepting user-supplied regex/grammar constraints is at risk of a low-effort denial of service. Patch to vLLM 0.24.0 immediately; until then, strip or hard-validate any user-controlled regex/grammar input before it reaches structured_outputs, and monitor for workers stuck at 100% CPU with no token throughput as a detection signal.

No confirmed active exploitation of CVE-2026-55574 has been reported, but organizations should still patch proactively.

Upgrade to vLLM 0.24.0 or later, which adds compilation timeout/complexity guards to the structured-output regex path. Until patched, do not expose the structured_outputs.regex parameter (or any wrapper API that maps to it) to untrusted or external users — restrict grammar/regex definition to server-side, developer-controlled schemas only. As a stopgap, front vLLM with a request-level timeout and per-worker circuit breaker that kills and restarts hung inference processes, and add input validation that rejects regex patterns with nested quantifiers or excessive repetition constructs before they reach the compiler. Detection: monitor for individual vLLM worker processes pegged at high CPU with no corresponding token generation progress, and alert on requests containing regex-like fields with unusually long or deeply nested patterns.

This vulnerability affects the following AI/ML architecture patterns: model serving, agent frameworks, RAG pipelines.

CVE-2026-55574 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.32%.