CVEReports
CVEReports

Automated vulnerability intelligence platform. Comprehensive reports for high-severity CVEs generated by AI.

Product

  • Home
  • Sitemap
  • RSS Feed

Company

  • About
  • Contact
  • Privacy Policy
  • Terms of Service

© 2026 CVEReports. All rights reserved.

Made with love by Amit Schendel & Alon Barad



CVE-2026-105753

CVE-2026-105753: Reachable Assertion in vLLM Multimodal IPC Cache Leading to Denial of Service

Alon Barad
Alon Barad
Software Engineer

Oct 6, 2026·5 min read·23 visits

Executive Summary (TL;DR)

A state desynchronization between vLLM's frontend and backend caches allows remote authenticated users to trigger a fatal assertion crash (CWE-617) by submitting a validation-failing multimodal request followed by a duplicate request, leading to complete server denial of service.

A state desynchronization (cache drift) vulnerability exists in the multimodal Inter-Process Communication (IPC) Least Recently Used (LRU) caches of vLLM. When a multimodal request fails validation after its media hash has been registered on the frontend but before the payload is committed to the backend engine core, the frontend and backend caches drift out of lockstep. A subsequent request reusing the same media triggers an assertion failure in the backend engine core, resulting in a complete denial of service.

Vulnerability Overview

vLLM uses a multi-process architecture to decouple request scheduling from high-performance GPU tensor computations. To optimize the processing of multimodal data such as images and video, vLLM implements a mirrored Inter-Process Communication (IPC) cache system.

This architecture consists of a frontend sender cache (MultiModalProcessorSenderCache on process P0) and an engine core receiver cache (MultiModalReceiverCache on process P1). The frontend cache acts as a shadow metadata tracker, while the backend cache stores the actual tensor payloads in GPU memory.

The system operates under a lockstep assumption. If the frontend registers a cache hit for a media hash, it transmits the request metadata with the payload data omitted (data=None), assuming the backend already possesses the cached tensor. A breakdown in this lockstep coordination exposes a reachable assertion bug class (CWE-617), allowing an attacker to terminate the backend engine process via network requests.

Root Cause Analysis

The root cause of the vulnerability is an unidirectional state mutation during the rendering phase that is not rolled back when downstream request validation fails.

When a multimodal request is processed, P0 renders the media, computes its unique cryptographic hash (mm_hash), and commits it directly to the P0 shadow metadata cache. If the request is subsequently rejected during validation (for example, if the prompt length exceeds the model's maximum limit and raises a ValueError), the request is discarded. Consequently, the actual tensor payload is never transmitted to or registered in P1's receiver cache.

This sequence causes state desynchronization: P0 assumes P1 has cached the media, whereas P1 has no record of the hash. When a subsequent request is submitted with the identical media, P0 detects a shadow cache hit and forwards the request with data=None. Upon receiving the request, P1 queries its receiver cache, returns None, and hits a fatal assertion block in vllm/multimodal/cache.py which terminates the process.

Code Analysis

In vulnerable versions of vLLM (prior to 0.28.0), the engine core unconditionally assumed cache consistency. The receiver cache lookup in vllm/multimodal/cache.py contained the following check:

# Vulnerable assertion in vllm/multimodal/cache.py
if (cached_item := self._cache.get(mm_hash)) is not None:
    return cached_item
 
assert mm_item is not None, f"Expected a cached item for {mm_hash=}"

If the frontend registered a shadow cache hit, mm_item was received as None. If mm_hash was missing from self._cache due to previous request validation failures, the assertion failed, causing the engine to abort execution.

The vulnerability was resolved by implementing a dual mitigation strategy. First, the fatal assertion was replaced with a retryable exception in PR #46747:

# Patched implementation with recovery exception
if (cached_item := self._cache.get(mm_hash)) is not None:
    return cached_item
 
if mm_item is None:
    raise MultiModalCacheMissError([mm_hash])

Second, immediate cache rollback was introduced in the OpenAI entrypoint (vllm/entrypoints/openai/chat_completion/serving.py) via PR #51897 to discard cache entries if a ValueError occurs during token validation:

# Patched OpenAI serving entrypoint
try:
    max_tokens = get_max_tokens(...
    )
except ValueError:
    for rendered_input in engine_inputs:
        if mm_hashes := cast(
            MultiModalHashes | None, rendered_input.get("mm_hashes")
        ):
            self.renderer.discard_mm_cache_entries(mm_hashes)
    raise

Exploitation Methodology

An attacker can systematically exploit this flaw through a two-step request sequence targeting any multimodal model served by a vulnerable vLLM instance.

First, the attacker submits a request containing an image payload paired with an excessively long prompt designed to trigger a validation failure (such as exceeding max_model_len). P0 computes the image hash, adds it to the shadow cache, and then the validation fails. The backend rejects the request, leaving its receiver cache empty.

Second, the attacker submits a valid request containing the exact same image. P0 identifies the hash in its shadow cache, sends data=None to the backend, and the backend crashes upon encountering the missing reference. This crash halts inference operations for all active clients on the shared server.

Impact Assessment

The execution of this vulnerability results in a complete denial of service of the vLLM inference server. Because vLLM is typically deployed as a shared daemon serving multiple downstream applications or clients, crashing the engine core process interrupts all active inference sessions.

This vulnerability has a CVSS 3.1 base score of 6.5 (Medium), reflecting the requirement for network access with low privileges (PR:L) and the resulting high availability impact (A:H). While confidentiality and integrity remain unaffected, the ease of triggering this state desynchronization makes it a highly effective vectors for targeted service disruptions.

Remediation and Mitigation

Remediation requires upgrading the vLLM installation to version 0.28.0 or later, which integrates both the self-healing fallback protocol and proactive validation rollback.

If upgrading is not immediately possible, organizations can apply the following workarounds:

  1. Disable multimodal IPC caching by adjusting the engine configuration parameters, which forces full payload delivery for every request and eliminates the lockstep dependency.

  2. Implement upstream validation of prompt token lengths at the API gateway or proxy layer. This ensures that validation-failing requests are dropped before they reach vLLM's processing pipeline, preventing the initial cache desynchronization.

Fix Analysis (2)

Technical Appendix

CVSS Score
6.5/ 10
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

Affected Systems

vLLM Multimodal Serving Engine

Affected Versions Detail

Product
Affected Versions
Fixed Version
vllm
vllm-project
< 0.28.00.28.0
AttributeDetail
CWE IDCWE-617 (Reachable Assertion)
Attack VectorNetwork (AV:N)
CVSS Score6.5
ImpactComplete Denial of Service (DoS)
Exploit StatusProof of Concept (PoC) / Unit Tests available
KEV StatusNot listed in CISA KEV

MITRE ATT&CK Mapping

T1499Endpoint Denial of Service
Impact
CWE-617
Reachable Assertion

The software contains an assertion statement that can be triggered by an input from an untrusted source, causing the application to terminate abnormally or crash, leading to a denial of service.

References & Sources

  • [1]GitHub Security Advisory GHSA-ph3r-5jfg-f84f
  • [2]NVD CVE-2026-105753 Detail
  • [3]vLLM Pull Request #46747
  • [4]vLLM Pull Request #51897

Attack Flow Diagram

Press enter or space to select a node. You can then use the arrow keys to move the node around. Press delete to remove it and escape to cancel.
Press enter or space to select an edge. You can then press delete to remove it or escape to cancel.

More Reports

•35 minutes ago•CVE-2026-102945
4.9

CVE-2026-102945: Unrecovered Goroutine Panic via Malformed NGAP PathSwitchRequest in Aether Project AMF

CVE-2026-102945 is a denial-of-service vulnerability affecting the Access and Mobility Management Function (AMF) of the Aether Project (omec-project/amf) prior to version 3.1.6. When processing an NGAP PathSwitchRequest message over the 3GPP N2 interface, the HandlePathSwitchRequest function fails to verify the byte slice lengths of extensible APER bit strings inside the UESecurityCapabilities Information Element before accessing index zero. Because the NGAP receive goroutine lacks a panic recovery handler, supplying a zero-length bit string triggers an out-of-bounds slice read panic that terminates the entire AMF daemon.

Amit Schendel
Amit Schendel
1 views•11 min read
•about 2 hours ago•CVE-2026-103025
8.7

CVE-2026-103025: Unbounded Memory Allocation and Decompression Bomb in bep/imagemeta PNG Parser

CVE-2026-103025 (GHSA-7hvw-3c2q-45qg) is a high-severity denial-of-service (DoS) vulnerability in github.com/bep/imagemeta prior to version 0.17.3. The PNG metadata parser fails to validate declared chunk lengths prior to heap buffer allocation (CWE-789) and inflates zlib-compressed zTXt IPTC chunks without an output size limit (CWE-409). Additionally, out-of-bounds slice indexing in the PNG zTXt handler and IPTC character set resolver allows remote attackers to trigger runtime panics.

Amit Schendel
Amit Schendel
1 views•1 min read
•about 3 hours ago•CVE-2026-103033
5.3

CVE-2026-103033: Pre-Verification Nil-Map Panic and Denial of Service in ALTCHA Go Library

CVE-2026-103033 (GHSA-gfwj-45pc-w6r5) is an unauthenticated pre-verification Denial of Service (DoS) vulnerability in the Go implementation of the ALTCHA bot-prevention library (github.com/altcha-org/altcha-lib-go). In versions 2.0.0 and 2.0.1, the internal function parseServerSignature conditionally initializes the ServerSignatureVerificationData.Extra map only when url.ParseQuery returns no error. When an attacker submits a payload containing both an unrecognized query parameter and malformed percent-encoding in the VerificationData field, url.ParseQuery returns partially parsed parameters alongside an error. Because execution continues without initializing the Extra map, the subsequent loop attempts to write unknown parameters into a nil map, triggering a fatal Go runtime panic before cryptographic signature verification occurs.

Alon Barad
Alon Barad
2 views•10 min read
•about 4 hours ago•CVE-2026-114516
9.8

CVE-2026-114516: Pre-Authentication Path Traversal and Arbitrary File Read in Etherpad

A pre-authentication path traversal vulnerability exists in Etherpad versions 1.8.14 prior to 3.3.3 within the `/static/*` plugin asset handler. In `src/node/utils/Minify.ts`, an unconditional backslash-to-forward-slash conversion is executed after pathname sanitization has already concluded. On POSIX systems, this transforms benign backslash characters into directory traversal sequences (`../`), allowing unauthenticated remote attackers to read arbitrary files from the host filesystem, extract administrative credentials from `settings.json`, and achieve remote code execution via the plugin installation interface.

Alon Barad
Alon Barad
3 views•10 min read
•about 5 hours ago•CVE-2026-128135
7.1

CVE-2026-128135: Sandbox Escape and Arbitrary File Overwrite via Hardlink Following in NLTK ZIP Extraction

A local sandbox escape vulnerability exists in the Natural Language Toolkit (NLTK) versions 3.9.4 through 3.10.3 within its ZIP archive extraction implementation. Due to a type mismatch in member resolution and a failure to inspect filesystem link counts (st_nlink > 1), the extraction routine follows pre-planted hardlinks inside the extraction root, allowing a local attacker to truncate and overwrite arbitrary out-of-root files writable by the executing process.

Alon Barad
Alon Barad
5 views•10 min read
•about 6 hours ago•CVE-2026-129304
5.4

CVE-2026-129304: Improper Input Validation and Redis Glob Injection in Guard Core

CVE-2026-129304 is an improper input validation and Redis glob pattern injection vulnerability in Guard Core versions 1.0.0 through 3.12.0. The header parsing routine _extract_from_forwarded_header() extracts values from the X-Forwarded-For HTTP header without verifying that the extracted substring is a valid IPv4 or IPv6 address. In deployments configured with trusted_proxies where the target header index is client-controllable, arbitrary strings containing Redis glob metacharacters reach BehaviorTracker and are interpolated directly into Redis key names and blocking KEYS queries. This allows unauthenticated attackers to aggregate cross-client behavioral metrics, trigger false-positive rate limits, cause denial of service via O(N) Redis keyspace scans, or raise unhandled ValueError exceptions during automated IP ban operations.

Amit Schendel
Amit Schendel
4 views•10 min read