CVE-2026-100651: vLLM: prompt-length bypass on disagg endpoint causes DoS
MEDIUMCISA: TRACK*vLLM before 0.29.0 does not validate prompt length on the disaggregated serving endpoint /inference/v1/generate when a multimodal 'features' payload is sent, so an authenticated low-privilege client can submit an overlong token_ids list and crash the inference worker. The impact is availability only (CVSS 6.5, no confidentiality or integrity loss), and exploitation is trivial over the network with low complexity. Likelihood is currently modest: EPSS is 0.00331 (top 76%), the CVE is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK*. Blast radius is notable, though, because vLLM has 95 downstream dependents and 134 other CVEs on record, so shared inference platforms are the main concern. Upgrade to 0.29.0 or later, and until then do not expose the disaggregated endpoint to untrusted clients.
What is the risk?
Medium. Exploitation needs only network reach and a low-privilege account, with no user interaction and no specialized AI knowledge. Exposure is limited to deployments that enable the disaggregated serving endpoint and run multimodal processors that skip prompt-length checks (for example Nemotron Parse, Whisper, FireRedLID). The impact is worker failure and denial of service. There is no data disclosure or code execution. Active exploitation is not observed (not in KEV, no public PoC, low EPSS), so this is a track-and-patch item rather than an emergency. Multi-tenant or shared inference services carry the highest business risk, since one tenant can disrupt the others.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| vLLM | pip | — | No patch |
Do you use vLLM? You're affected.
How severe is it?
What is the attack surface?
What should I do?
1 step1) Upgrade vLLM to 0.29.0 or later (the fixed release per GHSA-3mqx-f33v-vgp9). 2) If you cannot patch yet, block or restrict access to /inference/v1/generate at the reverse proxy or API gateway, and expose it only to trusted internal services. 3) Enforce authentication and per-client rate limits on all vLLM endpoints. 4) Add gateway-side limits on request body size and token_ids length relative to the model's max_model_len. 5) Run workers under a supervisor or orchestrator with automatic restart and health checks to limit downtime. 6) Detect by alerting on repeated worker crashes or restarts, tracebacks from the input-batch copy, and requests to the disaggregated endpoint with unusually large token_ids arrays.
What does CISA's SSVC say?
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:
Frequently Asked Questions
What is CVE-2026-100651?
vLLM before 0.29.0 does not validate prompt length on the disaggregated serving endpoint /inference/v1/generate when a multimodal 'features' payload is sent, so an authenticated low-privilege client can submit an overlong token_ids list and crash the inference worker. The impact is availability only (CVSS 6.5, no confidentiality or integrity loss), and exploitation is trivial over the network with low complexity. Likelihood is currently modest: EPSS is 0.00331 (top 76%), the CVE is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK*. Blast radius is notable, though, because vLLM has 95 downstream dependents and 134 other CVEs on record, so shared inference platforms are the main concern. Upgrade to 0.29.0 or later, and until then do not expose the disaggregated endpoint to untrusted clients.
Is CVE-2026-100651 actively exploited?
No confirmed active exploitation of CVE-2026-100651 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-100651?
1) Upgrade vLLM to 0.29.0 or later (the fixed release per GHSA-3mqx-f33v-vgp9). 2) If you cannot patch yet, block or restrict access to /inference/v1/generate at the reverse proxy or API gateway, and expose it only to trusted internal services. 3) Enforce authentication and per-client rate limits on all vLLM endpoints. 4) Add gateway-side limits on request body size and token_ids length relative to the model's max_model_len. 5) Run workers under a supervisor or orchestrator with automatic restart and health checks to limit downtime. 6) Detect by alerting on repeated worker crashes or restarts, tracebacks from the input-batch copy, and requests to the disaggregated endpoint with unusually large token_ids arrays.
What systems are affected by CVE-2026-100651?
This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference, disaggregated prefill/decode serving, multimodal pipelines, speech and document-parsing pipelines.
What is the CVSS score for CVE-2026-100651?
CVE-2026-100651 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.38%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0029Denial of AI ServiceAML.T0040AI Model Inference API AccessAML.T0049Exploit Public-Facing ApplicationCompliance Controls Affected
What are the technical details?
Original Advisory
vLLM before 0.29.0 fails to enforce decoder prompt-length validation on the disaggregated serving endpoint /inference/v1/generate. When the request contains a 'features' (multimodal) payload, vllm/entrypoints/serve/disagg/serving.py builds a multimodal EngineInput directly from the caller-supplied token_ids, and GenerateRequest.token_ids (vllm/entrypoints/serve/disagg/protocol.py) is not checked against model_config.max_model_len. For multimodal processors that report skip_prompt_length_check=True (for example Nemotron Parse, Whisper, and FireRedLID), InputProcessor._validate_prompt_len() returns immediately for both encoder and decoder prompts, so an overlong prompt becomes an EngineCoreRequest and reaches the worker input-batch copy into a fixed max_model_len-wide NumPy row. A client able to reach the endpoint on an affected model configuration can therefore submit an overlong token_ids list to trigger a worker failure and denial of service. Fixed in 0.29.0.
Exploitation Scenario
An attacker with a low-privilege account or API key on a shared vLLM inference service finds that the disaggregated /inference/v1/generate endpoint is reachable. They send a request with a 'features' multimodal payload for a Whisper or Nemotron Parse model and a token_ids list longer than max_model_len. Validation is skipped, the oversized prompt reaches the worker, and the copy into the fixed-width array fails and crashes the worker. The attacker repeats the request each time the worker restarts, keeping the AI service down for all tenants at almost no cost.
Weaknesses (CWE)
CWE-400 — Uncontrolled Resource Consumption: The product does not properly control the allocation and maintenance of a limited resource.
- [Architecture and Design] Design throttling mechanisms into the system architecture. The best protection is to limit the amount of resources that an unauthorized user can cause to be expended. A strong authentication and access control model will help prevent such attacks from occurring in the first place. The login application should be protected against DoS attacks as much as possible. Limiting the database access, perhaps by caching result sets, can help minimize the resources expended. To further limit the potential for a DoS attack, consider tracking the rate of requests received from users and blocking requests that exceed a defined rate threshold.
- [Architecture and Design] Mitigation of resource exhaustion attacks requires that the target system either: The first of these solutions is an issue in itself though, since it may allow attackers to prevent the use of the system by a particular valid user. If the attacker impersonates the valid user, they may be able to prevent the user from accessing the server in question. The second solution is simply difficult to effectively institute -- and even when properly done, it does not provide a full solution. It simply makes the attack require more resources on the part of the attacker. recognizes the attack and denies that user further access for a given amount of time, or uniformly throttles all requests in order to make it more difficult to consume resources more quickly than they can again be freed.
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:HReferences
Timeline
Related Vulnerabilities
CVE-2026-6173210.0Decepticon: ChatML token injection yields sandbox RCE
Same package: vllmCVE-2024-110419.8vllm: RCE via unsafe pickle deserialization in MessageQueue
Same package: vllmCVE-2025-472779.8vLLM: RCE via exposed TCPStore in distributed inference
Same package: vllmCVE-2026-259609.8vllm: SSRF allows internal network access
Same package: vllmCVE-2024-90539.8vllm: RCE via unsafe pickle deserialization in RPC server
Same package: vllm