CVE-2026-100654: vLLM: out-of-range stop_token_ids crash EngineCore (DoS)
MEDIUMvLLM before 0.29.0 does not check that user-supplied stop_token_ids fall within the model's vocabulary, so a single malformed request to /v1/completions or /v1/chat/completions with min_tokens > 0 triggers a CUDA device-side assertion that leaves the inference engine in a fatal state until it is restarted. Any authenticated API user, including a low-privilege tenant or a leaked key, can take down a shared GPU inference service with one request. Exploitation likelihood is modest, with an EPSS of 0.003 (top 78% percentile), no public exploit, no Nuclei template, no KEV listing and a CISA SSVC decision of TRACK. Still, the trigger is trivial and vLLM has 95 downstream dependents and 134 other CVEs on record, so the availability risk is real for multi-tenant or internet-exposed deployments. Upgrade to vLLM 0.29.0 or later, and until then reject or clamp stop_token_ids at a gateway and make sure the service is auto-restarted by your orchestrator.
What is the risk?
Medium (CVSS 3.1: 6.5, AV:N/AC:L/PR:L/UI:N, availability impact only). Exploitation is trivial for anyone holding a valid API credential, but there is no confidentiality or integrity impact, no known exploitation in the wild, no public exploit and a low EPSS (0.00307). Risk rises for shared, multi-tenant or internet-facing inference endpoints where a single request can cause an outage affecting all users, and where GPU capacity is costly to recover.
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 (check the GHSA-v5gm-qgmv-gc6c advisory for the fixed release). 2) Until patched, add validation at an API gateway or reverse proxy to reject requests whose stop_token_ids contain negative values or values >= the model vocabulary size, or strip stop_token_ids entirely for untrusted clients. 3) Restrict API access with strong authentication, per-key rate limits and network segmentation; rotate any exposed keys. 4) Configure liveness and readiness probes that detect a dead EngineCore and restart the container automatically. 5) Detect by alerting on HTTP 500 spikes on /v1/completions and /v1/chat/completions, CUDA 'device-side assert triggered' messages in vLLM logs, and unexpected engine restarts or requests carrying unusually large stop_token_ids values.
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-2026-100654?
vLLM before 0.29.0 does not check that user-supplied stop_token_ids fall within the model's vocabulary, so a single malformed request to /v1/completions or /v1/chat/completions with min_tokens > 0 triggers a CUDA device-side assertion that leaves the inference engine in a fatal state until it is restarted. Any authenticated API user, including a low-privilege tenant or a leaked key, can take down a shared GPU inference service with one request. Exploitation likelihood is modest, with an EPSS of 0.003 (top 78% percentile), no public exploit, no Nuclei template, no KEV listing and a CISA SSVC decision of TRACK. Still, the trigger is trivial and vLLM has 95 downstream dependents and 134 other CVEs on record, so the availability risk is real for multi-tenant or internet-exposed deployments. Upgrade to vLLM 0.29.0 or later, and until then reject or clamp stop_token_ids at a gateway and make sure the service is auto-restarted by your orchestrator.
Is CVE-2026-100654 actively exploited?
No confirmed active exploitation of CVE-2026-100654 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-100654?
1) Upgrade vLLM to 0.29.0 or later (check the GHSA-v5gm-qgmv-gc6c advisory for the fixed release). 2) Until patched, add validation at an API gateway or reverse proxy to reject requests whose stop_token_ids contain negative values or values >= the model vocabulary size, or strip stop_token_ids entirely for untrusted clients. 3) Restrict API access with strong authentication, per-key rate limits and network segmentation; rotate any exposed keys. 4) Configure liveness and readiness probes that detect a dead EngineCore and restart the container automatically. 5) Detect by alerting on HTTP 500 spikes on /v1/completions and /v1/chat/completions, CUDA 'device-side assert triggered' messages in vLLM logs, and unexpected engine restarts or requests carrying unusually large stop_token_ids values.
What systems are affected by CVE-2026-100654?
This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference endpoints, OpenAI-compatible API gateways, RAG pipelines, agent frameworks.
What is the CVSS score for CVE-2026-100654?
CVE-2026-100654 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.31%.
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 accepts user-controlled stop_token_ids on the OpenAI-compatible POST /v1/completions and POST /v1/chat/completions endpoints but validates only that the values are integers, not that each token id is within the model vocabulary/logits range. When min_tokens > 0, the stop token ids are used as logits indices to suppress stop tokens, so an out-of-range id reaches a CUDA indexing operation (index_put_) and triggers a device-side assertion. An authenticated API user can send a single malformed completion request that returns 500 Internal Server Error and puts EngineCore into a fatal state, causing subsequent requests to fail until the service is restarted (denial of service).
Exploitation Scenario
An attacker with a low-privilege API key to a shared vLLM-backed internal LLM gateway, for example a compromised developer token or a free-tier tenant, sends a single POST to /v1/chat/completions with min_tokens set to 1 and stop_token_ids containing an integer far above the vocabulary size. The server returns a 500 and the CUDA assertion leaves EngineCore broken. All other tenants' requests then fail until an operator restarts the service, and the attacker can repeat the request after each restart to sustain the outage.
Weaknesses (CWE)
CWE-129 — Improper Validation of Array Index: The product uses untrusted input when calculating or using an array index, but the product does not validate or incorrectly validates the index to ensure the index references a valid position within the array.
- [Architecture and Design] Use an input validation framework such as Struts or the OWASP ESAPI Validation API. Note that using a framework does not automatically address all input validation problems; be mindful of weaknesses that could arise from misusing the framework itself (CWE-1173).
- [Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server. Even though client-side checks provide minimal benefits with respect to server-side security, they are still useful. First, they can support intrusion detection. If the server receives input that should have been rejected by the client, then it may be an indication of an attack. Second, client-side error-checking can provide helpful feedback to the user about the expectations for valid input. Third, there may be a reduction in server-side processing time for accidental input errors, although this is typically a small savings.
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