CVE-2026-100647: vLLM: unbounded cache_salt allows unauthenticated DoS

HIGH
Published September 26, 2026
CISO Take

vLLM versions before 0.29.0 accept an unbounded cache_salt parameter on their OpenAI-compatible and Anthropic API endpoints. An unauthenticated attacker can send a request with a multi-hundred-megabyte salt, which forces expensive pickle serialization and SHA-256 hashing on the single EngineCore scheduler thread and stalls inference for every concurrent user. It is network-reachable, needs no privileges or user interaction, and carries a CVSS of 7.5, but it only affects availability. Exploitation likelihood is modest: EPSS is 0.00348 (top 74th percentile), there is no public exploit or Nuclei template, it is not in CISA KEV, and CISA SSVC says TRACK. The blast radius is still real, with 95 downstream dependents and 134 other CVEs in the same package, which signals heavy scrutiny and frequent patching. Upgrade to vLLM 0.29.0 or later, and until then cap request body size at the reverse proxy or gateway and keep inference endpoints off the public internet.

Sources:NVDEPSSGitHub AdvisoryVulnCheckCISA SSVCATLASOpenSSF

What is the risk?

Medium-high for internet-exposed or multi-tenant vLLM deployments, and low-to-moderate for properly isolated internal ones. The attack is trivial to execute (AV:N/AC:L/PR:N/UI:N) and a single request can stall the whole engine. Impact is limited to availability, with no confidentiality or integrity loss. Likelihood of exploitation is currently low: EPSS 0.00348, no public PoC, no scanner template, not in KEV, SSVC TRACK. Risk rises sharply for production LLM APIs that sit directly on the internet without a body-size-limiting proxy or authentication, because one scheduler thread serves all tenants.

How does the attack unfold?

Reconnaissance and entry
Attacker identifies an exposed vLLM OpenAI- or Anthropic-compatible endpoint that requires no authentication.
AML.T0049
Exploitation
Attacker sends a request with a multi-hundred-megabyte cache_salt that triggers pickle serialization and SHA-256 hashing on the EngineCore scheduler thread.
AML.T0034.001
Impact
The scheduler stalls and all concurrent inference requests time out, producing a service-wide outage that repeated requests can sustain.
AML.T0029

What systems are affected?

PackageEcosystemVulnerable RangePatched
vLLMpip—No patch
96.9K93 dependentsPushed 3d ago26% patched~37d to patchFull package profile →

Do you use vLLM? You're affected.

How severe is it?

CVSS 3.1
7.5 / 10
EPSS
0.3%
chance of exploitation in 30 days
Higher than 29% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Trivial

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRNone
UINone
SUnchanged
CNone
INone
AHigh

What should I do?

1 step
  1. 1) Upgrade vLLM to 0.29.0 or later, which adds validation on cache_salt (patched version taken from the advisory title; confirm against the GHSA). 2) Until patched, enforce a maximum request body size (for example a few MB) at the reverse proxy, API gateway or WAF in front of vLLM, and reject or strip cache_salt on untrusted traffic. 3) Require authentication (API keys, mTLS or a gateway) and do not expose vLLM ports directly to the internet. 4) Add per-client rate limiting and connection limits. 5) Detect by alerting on unusually large request bodies, sudden scheduler stalls or queue growth, high CPU on the EngineCore process, and failed health checks, and by reviewing access logs for requests containing very large cache_salt fields. 6) Use autoscaling and fast restart policies to limit blast radius, and isolate tenants onto separate instances where feasible.

What does CISA's SSVC say?

DecisionTrack
Exploitationnone
AutomatableYes
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.6.2.6-AI system operation and monitoring
NIST AI RMF
MEASURE 2.7-AI system security and resilience are evaluated and documented
OWASP LLM Top 10
LLM10:2025-Unbounded Consumption

Frequently Asked Questions

What is CVE-2026-100647?

vLLM versions before 0.29.0 accept an unbounded cache_salt parameter on their OpenAI-compatible and Anthropic API endpoints. An unauthenticated attacker can send a request with a multi-hundred-megabyte salt, which forces expensive pickle serialization and SHA-256 hashing on the single EngineCore scheduler thread and stalls inference for every concurrent user. It is network-reachable, needs no privileges or user interaction, and carries a CVSS of 7.5, but it only affects availability. Exploitation likelihood is modest: EPSS is 0.00348 (top 74th percentile), there is no public exploit or Nuclei template, it is not in CISA KEV, and CISA SSVC says TRACK. The blast radius is still real, with 95 downstream dependents and 134 other CVEs in the same package, which signals heavy scrutiny and frequent patching. Upgrade to vLLM 0.29.0 or later, and until then cap request body size at the reverse proxy or gateway and keep inference endpoints off the public internet.

Is CVE-2026-100647 actively exploited?

No confirmed active exploitation of CVE-2026-100647 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-100647?

1) Upgrade vLLM to 0.29.0 or later, which adds validation on cache_salt (patched version taken from the advisory title; confirm against the GHSA). 2) Until patched, enforce a maximum request body size (for example a few MB) at the reverse proxy, API gateway or WAF in front of vLLM, and reject or strip cache_salt on untrusted traffic. 3) Require authentication (API keys, mTLS or a gateway) and do not expose vLLM ports directly to the internet. 4) Add per-client rate limiting and connection limits. 5) Detect by alerting on unusually large request bodies, sudden scheduler stalls or queue growth, high CPU on the EngineCore process, and failed health checks, and by reviewing access logs for requests containing very large cache_salt fields. 6) Use autoscaling and fast restart policies to limit blast radius, and isolate tenants onto separate instances where feasible.

What systems are affected by CVE-2026-100647?

This vulnerability affects the following AI/ML architecture patterns: Model serving / LLM inference, OpenAI-compatible API gateways, Multi-tenant LLM platforms, RAG pipelines, Agent frameworks.

What is the CVSS score for CVE-2026-100647?

CVE-2026-100647 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.35%.

What is the AI security impact?

Affected AI Architectures

Model serving / LLM inferenceOpenAI-compatible API gatewaysMulti-tenant LLM platformsRAG pipelinesAgent frameworks

MITRE ATLAS Techniques

AML.T0029Denial of AI Service
AML.T0034.001Resource-Intensive Queries
AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15
ISO 42001: A.6.2.6
NIST AI RMF: MEASURE 2.7
OWASP LLM Top 10: LLM10:2025

What are the technical details?

Original Advisory

vLLM versions before 0.29.0 contain a denial-of-service vulnerability in the cache_salt parameter accepted on OpenAI-compatible and Anthropic API endpoints, which lacks maximum length validation and is processed on the single EngineCore scheduler thread. Unauthenticated attackers can send HTTP requests with multi-hundred-megabyte salt values that trigger expensive pickle serialization and SHA-256 hashing, stalling the scheduler thread and denying service to all concurrent requests.

Exploitation Scenario

An attacker finds a company's public LLM gateway that fronts vLLM through an OpenAI-compatible /v1/chat/completions endpoint. They send a few HTTP POST requests whose cache_salt field holds a few hundred megabytes of data. The vLLM server pickles and SHA-256 hashes the value on the single EngineCore scheduler thread, so the scheduler stalls and every other tenant's requests time out. The attacker repeats the request at low rate to keep the service down, causing an outage of the customer-facing AI assistant and downstream agent workflows. No credentials or model knowledge are needed.

Weaknesses (CWE)

CWE-20 — Improper Input Validation: The product receives input or data, but it does not validate or incorrectly validates that the input has the properties that are required to process the data safely and correctly.

  • [Architecture and Design] Consider using language-theoretic security (LangSec) techniques that characterize inputs using a formal language and build "recognizers" for that language. This effectively requires parsing to be a distinct layer that effectively enforces a boundary between raw input and internal data representations, instead of allowing parser code to be scattered throughout the program, where it could be subject to errors or inconsistencies that create weaknesses. [REF-1109] [REF-1110] [REF-1111]
  • [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).

Source: MITRE CWE corpus.

CVSS Vector

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

Timeline

Published
September 26, 2026
Last Modified
October 6, 2026
First Seen
September 26, 2026

Related Vulnerabilities