CVE-2026-73559
Received Received - Intake

Resource Exhaustion in vLLM API Completions

Publication date: 2026-08-13

Last updated on: 2026-08-13

Assigner: GitHub, Inc.

Description
vLLM is an inference and serving engine for large language models. From 0.19.0 until 0.26.0, the /v1/completions CompletionRequest.prompt field in vllm/entrypoints/openai/completion/protocol.py accepts an unbounded list[str] or list[list[int]], prompt_to_seq() in vllm/renderers/inputs/preprocess.py and OnlineRenderer.preprocess_completion() in vllm/renderers/online_renderer.py expand every element, and vllm/entrypoints/openai/completion/serving.py creates one engine generator and response slot per prompt, allowing an authenticated API client to exhaust CPU, memory, async scheduling capacity, engine request slots, and response buffering with one request. This issue is fixed in version 0.26.0.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published
2026-08-13
Last Modified
2026-08-13
Generated
2026-08-13
AI Q&A
2026-08-13
EPSS Evaluated
N/A
NVD
CVE-2026-73559
EUVD
EUVD-2026-58067

Affected Vendors & Products

Showing 2 associated CPEs
Vendor Product Version / Range
vllm_project vllm From 0.19.0 (inc) to 0.26.0 (inc)
vllm_project vllm 0.26.0

Helpful Resources

Exploitability

CWE
CWE Icon
KEV
KEV Icon
CWE ID Description
CWE-400 The product does not properly control the allocation and maintenance of a limited resource.

Attack-Flow Graph

AI Quick Actions

Instant insights powered by AI
Executive Summary

This vulnerability in vLLM (versions 0.19.0 to 0.26.0) allows an authenticated API client to send a prompt as an unbounded list of strings or token IDs to the /v1/completions endpoint. Each element in the list triggers a separate backend engine request, creating one generator and response slot per prompt. This can exhaust CPU, memory, scheduling capacity, and response buffering, leading to denial-of-service conditions for other users.

Detection Guidance

To detect this vulnerability, monitor for unusually large prompt lists in /v1/completions requests. Check logs for requests exceeding 1024 prompts or excessive resource usage during API calls. Use network monitoring tools to identify repeated or malformed requests targeting the endpoint.

Impact Analysis

An attacker could send a single malicious request with a very large prompt list, causing the server to allocate excessive resources. This may lead to degraded performance, crashes, or unavailability for other legitimate users sharing the same system. The impact includes CPU exhaustion, memory overload, and scheduling delays.

Compliance Impact

This vulnerability could lead to resource exhaustion attacks, causing system downtime or degraded performance. For GDPR, this may impact availability of services processing personal data, potentially violating Article 32 requirements for security and integrity. For HIPAA, it could disrupt access to protected health information systems, affecting the availability requirement under the Security Rule.

Mitigation Strategies

Upgrade vLLM to version 0.26.0 or later. Set the VLLM_MAX_COMPLETION_PROMPTS environment variable to a lower value if needed. Implement rate limiting and input validation for the /v1/completions endpoint to reject oversized prompt lists before processing.

Chat Assistant

Ask questions about this CVE
Hi! I’m here to help you understand CVE-2026-73559. Ask me anything about the vulnerability, its impact, or mitigation strategies.
0/70

EPSS Chart