CVE-2025-48943: vLLM: ReDoS crashes inference server via malformed regex

GHSA-9hcf-v7m4-6m2jMEDIUMPoC AVAILABLE
Published May 30, 2025
CISO Take

Any authenticated user of vLLM 0.8.x can crash the entire inference server by submitting a malformed regex as a structured output constraint — no special skills required. This is a shared-resource risk: one bad request takes down inference for all downstream services and agents. Upgrade to vLLM 0.9.0 immediately; if blocked, add regex validation at the API gateway layer.

What is the risk?

Operationally higher risk than CVSS 6.5 suggests for shared inference infrastructure. Low Complexity + Low Privileges means any authenticated API consumer — including internal developers, CI pipelines, or compromised service accounts — can trigger a full server crash with a single request. EPSS is negligible (0.00084), indicating no observed exploitation yet, but the attack is trivially reproducible once the advisory is public. The sibling vulnerability CVE-2025-48942 (same pattern, JSON schema instead of regex) confirms a systemic input validation gap in vLLM's structured output subsystem.

What systems are affected?

PackageEcosystemVulnerable RangePatched
vLLMpip—No patch
96.9K93 dependentsPushed 3d ago26% patched~37d to patchFull package profile →
vLLMpip>= 0.8.0, < 0.9.00.9.0
96.9K93 dependentsPushed 3d ago26% patched~37d to patchFull package profile →

How severe is it?

CVSS 3.1
6.5 / 10
EPSS
0.5%
chance of exploitation in 30 days
Higher than 44% of all CVEs
Exploitation Status
Exploit Available
Exploitation: MEDIUM
Sophistication
Trivial
Exploitation Confidence
medium
○Public PoC indexed (trickest/cve)
Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRLow
UINone
SUnchanged
CNone
INone
AHigh

What should I do?

6 steps
  1. PATCH

    Upgrade vLLM to >= 0.9.0 immediately (pip install --upgrade vllm).

  2. WORKAROUND

    If upgrade is blocked, add a pre-validation layer that sanitizes regex patterns before forwarding to vLLM — reject patterns exceeding complexity thresholds (e.g., nested quantifiers).

  3. RATE-LIMIT: Apply per-user rate limiting on structured output endpoints to slow down brute-force crash attempts.

  4. DETECT

    Monitor for sudden vLLM process restarts or spikes in 5xx errors on structured output endpoints. Alert on repeated server crashes from the same API key/user.

  5. ISOLATE

    Run vLLM behind an internal-only API gateway; avoid exposing the inference API directly to untrusted users.

  6. AUDIT

    Review whether structured output endpoints are exposed to external or low-trust identities.

What does CISA's SSVC say?

DecisionTrack
Exploitationnone
AutomatableNo
Technical Impactpartial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Article 15-Accuracy, robustness and cybersecurity
ISO 42001
A.6.2.8-AI system availability and performance
NIST AI RMF
GOVERN-1.7-Processes for AI risk identification and responseMANAGE-2.2-Mechanisms to sustain AI system value and manage risks over time
OWASP LLM Top 10
LLM04-Model Denial of Service

Frequently Asked Questions

What is CVE-2025-48943?

Any authenticated user of vLLM 0.8.x can crash the entire inference server by submitting a malformed regex as a structured output constraint — no special skills required. This is a shared-resource risk: one bad request takes down inference for all downstream services and agents. Upgrade to vLLM 0.9.0 immediately; if blocked, add regex validation at the API gateway layer.

Is CVE-2025-48943 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2025-48943, increasing the risk of exploitation.

How to fix CVE-2025-48943?

1. PATCH: Upgrade vLLM to >= 0.9.0 immediately (pip install --upgrade vllm). 2. WORKAROUND: If upgrade is blocked, add a pre-validation layer that sanitizes regex patterns before forwarding to vLLM — reject patterns exceeding complexity thresholds (e.g., nested quantifiers). 3. RATE-LIMIT: Apply per-user rate limiting on structured output endpoints to slow down brute-force crash attempts. 4. DETECT: Monitor for sudden vLLM process restarts or spikes in 5xx errors on structured output endpoints. Alert on repeated server crashes from the same API key/user. 5. ISOLATE: Run vLLM behind an internal-only API gateway; avoid exposing the inference API directly to untrusted users. 6. AUDIT: Review whether structured output endpoints are exposed to external or low-trust identities.

What systems are affected by CVE-2025-48943?

This vulnerability affects the following AI/ML architecture patterns: LLM inference serving, structured output pipelines, AI agent frameworks, multi-tenant LLM API gateways, RAG pipelines with constrained generation.

What is the CVSS score for CVE-2025-48943?

CVE-2025-48943 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.51%.

What is the AI security impact?

Affected AI Architectures

LLM inference servingstructured output pipelinesAI agent frameworksmulti-tenant LLM API gatewaysRAG pipelines with constrained generation

MITRE ATLAS Techniques

AML.T0029Denial of AI Service
AML.T0034Cost Harvesting
AML.T0040AI Model Inference API Access
AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15
ISO 42001: A.6.2.8
NIST AI RMF: GOVERN-1.7, MANAGE-2.2
OWASP LLM Top 10: LLM04

What are the technical details?

Original Advisory

vLLM is an inference and serving engine for large language models (LLMs). Version 0.8.0 up to but excluding 0.9.0 have a Denial of Service (ReDoS) that causes the vLLM server to crash if an invalid regex was provided while using structured output. This vulnerability is similar to GHSA-6qc9-v4r8-22xg/CVE-2025-48942, but for regex instead of a JSON schema. Version 0.9.0 fixes the issue.

Exploitation Scenario

An adversary with low-privilege API access (e.g., a developer API key, a compromised service account, or a malicious internal user) sends a POST request to the vLLM inference endpoint with a carefully crafted regex pattern in the guided generation parameters — such as a pattern with catastrophic backtracking like `(a+)+$`. The vLLM server attempts to compile and validate the regex, triggering exponential backtracking that consumes all CPU and crashes the process. All concurrent inference requests fail. In a Kubernetes environment without proper liveness probes, the pod may hang rather than restart, causing a prolonged outage. An adversary can repeat this attack to maintain denial of service against a critical LLM serving layer.

Weaknesses (CWE)

CWE-248 — Uncaught Exception: An exception is thrown from a function, but it is not caught.

  • [Requirements] The choice between a language which has named or unnamed exceptions needs to be done. While unnamed exceptions exacerbate the chance of not properly dealing with an exception, named exceptions suffer from the up call version of the weak base class problem.
  • [Requirements] A language can be used which requires, at compile time, to catch all serious exceptions. However, one must make sure to use the most current version of the API as new exceptions could be added.

Source: MITRE CWE corpus.

CVSS Vector

CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

Timeline

Published
May 30, 2025
Last Modified
June 27, 2025
First Seen
May 30, 2025

Related Vulnerabilities