Zaloguj się
Deutsch English Español Polski
Średnia

vLLM's Artifact Pin Decay allows pinned deployments to load unpinned code, weights, and processors

  • Ujawniono: 10.06.2026
  • Aktualizacja: 11.10.2026 15:17

Summary

vLLM's revision pinning controls do not consistently apply to all artifacts loaded for a model. A deployment that supplies --revision or --code-revision can still load dynamic code, GGUF files, image processors, retrieval side weights, or same-repository subfolder weights/config from an unpinned/default revision.

This is a supply-chain integrity issue for pinned vLLM deployments. Operators can believe they are serving a reviewed model revision while vLLM resolves behavior-affecting nested or sibling artifacts outside that reviewed revision.

Details

The expected invariant is:

When a vLLM operator supplies a model or code revision pin, every code, config, processor, weight file, side weight, and same-repository subfolder artifact loaded as part of that model should resolve under that pin unless vLLM exposes and enforces a separate explicit pin for that artifact.

Current main was verified affected at commit 3795d7acf431980e62e738493f437ae2a51549da.

Affected source boundaries:

  • vllm/model_executor/models/registry.py:1045-1051 and :1058-1064
    • _try_resolve_transformers() passes revision=model_config.revision and trust_remote_code=model_config.trust_remote_code, but omits code_revision=model_config.code_revision for external auto_map dynamic module imports.
  • vllm/model_executor/model_loader/gguf_loader.py:58-60
    • The direct-file GGUF form repo/file.gguf calls hf_hub_download(repo_id=repo_id, filename=filename) without passing revision.
  • vllm/model_executor/models/roberta.py:203-209
    • BGE-M3 secondary sparse and ColBERT side weights are declared with revision=None.
  • vllm/model_executor/models/kimi_k25.py:111-114
    • Kimi-K2.5 calls cached_get_image_processor() without passing model_config.revision.
  • vllm/model_executor/models/kimi_audio.py:92-95
    • Kimi-Audio loads Whisper config from the whisper-large-v3 subfolder without a revision argument.
  • vllm/model_executor/models/kimi_audio.py:425-430
    • Kimi-Audio declares same-repository whisper-large-v3 secondary weights with revision=None.
  • vllm/model_executor/model_loader/default_loader.py:287-301
    • The default loader preserves model_config.revision for the primary source, then consumes model-supplied secondary sources as declared.

The strongest example is Kimi-Audio: the primary moonshotai/Kimi-Audio-7B-Instruct weights preserve the configured model revision, but the same-repository whisper-large-v3 audio tower config/weights do not. A pinned Kimi-Audio deployment can therefore load the Whisper subfolder outside the audited revision.

This report does not claim a trust_remote_code=False bypass, unauthenticated RCE, or real artifact compromise. The issue is improper propagation of explicit artifact pins across supported loader paths.

Impact

Affected users are operators who pin vLLM model deployments to a reviewed Hugging Face revision for safety review, provenance, rollback, or reproducibility. The impact is that the pin does not reliably describe the full set of artifacts vLLM serves. Even when the operator selects an audited revision, vLLM can resolve behavior-affecting secondary artifacts from the repository default branch or another mutable ref.

Depending on the model path, the unpinned artifact can be dynamic model code, a GGUF file, an image processor, retrieval side weights, or the same-repository Kimi-Audio Whisper subfolder weights/config.

This breaks the operational guarantee of a pinned deployment: "serve the exact artifact set I reviewed." A later change to an unpinned secondary artifact can alter model behavior without changing the operator's configured revision, making review, rollback, incident response, and audit records unreliable.

Occurrences

  • vllm/model_executor/models/kimi_k25.py L111-L114 — Kimi-K2.5 loads its image processor with cached_get_image_processor() but does not pass self.ctx.model_config.revision. The processor can therefore resolve from the default repository revision even when the model deployment is pinned.
  • vllm/model_executor/models/kimi_audio.py L425-L430 — Kimi-Audio declares same-repository whisper-large-v3 secondary weights with revision=None. A pinned Kimi-Audio deployment can therefore load the Whisper audio tower weights from an unpinned/default revision.
  • vllm/model_executor/models/kimi_audio.py L92-L95 — Kimi-Audio loads Whisper config from the same repository's whisper-large-v3 subfolder without passing the top-level model revision. The config for this behavior-affecting subcomponent can be resolved outside the audited model revision.
  • vllm/model_executor/models/registry.py L1058-L1064 — The later dynamic model-class resolution repeats the same pin-decay pattern: it forwards revision and trust_remote_code, but omits code_revision. This means an operator-provided code pin is not enforced at the dynamic module loader boundary.
  • vllm/model_executor/model_loader/gguf_loader.py L58-L60 — The direct GGUF form repo/file.gguf calls hf_hub_download(repo_id=repo_id, filename=filename) without passing model_config.revision. A deployment that pins the model revision can therefore resolve this GGUF file from the repository default revision.
  • vllm/model_executor/models/registry.py L1045-L1051 — try_get_class_from_dynamic_module() is called for external auto_map config/model classes with revision=model_config.revision, but without forwarding model_config.code_revision. When --code-revision is set, this dynamic module resolution can still fall back to the default code revision instead of the audited code revision.
  • vllm/model_executor/models/roberta.py L203-L209 — BgeM3EmbeddingModel creates same-repository secondary sparse/ColBERT weight sources with revision=None. The primary model revision is not propagated to these side weights, so they can be downloaded outside the operator-selected model revision.

Fixes

This was fixed in: https://github.com/vllm-project/vllm/pull/42616

Originally filed via huntr: https://huntr.com/bounties/3f1e24c0-87d2-4f6c-a705-820f380879ac.

The vLLM maintainer (Russell Bryant) redirected the report to the private GHSA channel. Offline proof bundle (vllm_artifact_pin_decay_bundle_verify.py + bundle-verification-20260430T143506Z.json) is available upon request.

Czego dotyczy

Kliknij technologię, aby zobaczyć wszystkie dotyczące jej podatności.

PlatformaPaczkaPodatne wersjePoprawka w wersji
Python (PyPI) vllm < 0.22.0 0.22.0

Odnośniki

Ocena CVSS 6.5/10
Wektor CVSS
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:H/A:N
Słabość (CWE)
CWE-345
Platforma
Python (PyPI)
Paczka
vllm
Ujawniono
10.06.2026
Aktualizacja
11.10.2026 15:17

Źródła i licencje

OSV / GitHub Advisory Database - PYSEC-2026-2301 - PyPA Advisory Database; CC BY 4.0

OSV / GitHub Advisory Database - GHSA-3ww4-5jv9-j5gm - CC BY 4.0

Historia podatności: vllm

Wszystkie opisane podatności tej paczki (Python (PyPI)), od najnowszej. Widać, czy to jednorazowa wpadka, czy paczka, której trzeba pilnować.

Wszystkie (94) w archiwum
Wszystkich wpisów
94
Pierwsza → ostatnia
09.2024 10.2026
Bez poprawki
5
Wykorzystywana w atakach
0

Rozkład istotności

  • Krytyczna 9
  • Wysoka 23
  • Średnia 47
  • Niska 4
  • Nieoceniona 11

Ujawnienia w latach

Istotność:
  1. A flaw has been found in vllm-project vLLM up to 0.26.0. This vulnerability affects unknown code of the file rust/src/parser/src/unified/gemma4.rs of the component Gemma4UnifiedParser. Executing a man
    Wysoka · 7.5 CVE-2026-103241 < 0.27.0 0.27.0
  2. vLLM versions 0.22.0 through 0.23.0 fail to validate stop_token_ids against vocabulary bounds in Rust HTTP and gRPC frontends, allowing out-of-vocabulary token IDs to reach MinTokensLogitsProcessor. A
    Wysoka · 7.5 CVE-2026-100652 >= 0.22.0, < 0.24.0 0.24.0
  3. vLLM has Remote DoS via Invalid Recovered Token Reinjection
    Wysoka · 7.5 CVE-2026-54234 >= 0.17.1, < 0.24.0 0.24.0
  4. vLLM: ReDoS via structured_outputs.regex compiled without timeout in xgrammar and outlines backends
    Wysoka · 7.5 CVE-2026-55574 < 0.24.0 0.24.0
  5. vLLM denial of service via prompt embeds on M-RoPE models
    Wysoka CVE-2026-55514 >= 0.12.0, < 0.24.0 0.24.0
  6. vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.22.1, the vLLM Dockerfile is vulnerable to a dependency confusion attack through the flashinfer-jit-cache package.
    Wysoka · 8.8 CVE-2026-54232 < 0.22.1 0.22.1
  7. vLLM: OpenAI auth bypass
    Krytyczna · 9.1 CVE-2026-48746 >= 0.3.0, < 0.22.0 0.22.0
  8. vLLM: Security Check Bypass via assert Statement in Activation Function Loading Allows Arbitrary Code Execution
    Wysoka · 7.5 CVE-2026-41523 < 0.22.0 0.22.0
  9. vLLM versions 0.8.0 and later are vulnerable to an Out-of-Memory (OOM) Denial of Service (DoS) attack due to unbounded frame count processing in the `VideoMediaIO.load_base64()` method. When processin
    Wysoka · 7.5 CVE-2026-5497 >= 0.8.0, < 0.19.0 0.19.0
  10. vLLM's Artifact Pin Decay allows pinned deployments to load unpinned code, weights, and processors ten wpis
    Średnia · 6.5 CVE-2026-47155 < 0.22.0 0.22.0
  11. vLLM has Hardcoded Trust Override in Model Files Enables RCE Despite Explicit User Opt-Out
    Wysoka · 8.8 CVE-2026-27893 >= 0.10.1, < 0.18.0 0.18.0
  12. vLLM has RCE In Video Processing
    Krytyczna · 9.8 CVE-2026-22778 >= 0.8.3, < 0.14.1 0.14.1
  13. vLLM vulnerable to Server-Side Request Forgery (SSRF) through MediaConnector
    Wysoka · 7.1 CVE-2026-24779 < 0.14.1 0.14.1
  14. vLLM affected by RCE via auto_map dynamic module loading during model initialization
    Wysoka · 8.8 CVE-2026-22807 >= 0.10.1, < 0.14.0 0.14.0
  15. vLLM introduced enhanced protection for CVE-2025-62164
    Wysoka · 8.8 CVE-2026-56340 >= 0.10.2, < 0.13.0 0.13.0
  16. vLLM vulnerable to remote code execution via transformers_utils/get_config
    Wysoka · 7.1 CVE-2025-66448 < 0.11.1 0.11.1
  17. vLLM vulnerable to DoS with incorrect shape of multimodal embedding inputs
    Wysoka · 6.5 CVE-2025-62372 >= 0.5.5, < 0.11.1 0.11.1
  18. vLLM deserialization vulnerability leading to DoS and potential RCE
    Wysoka · 8.8 CVE-2025-62164 >= 0.10.2, < 0.11.1 0.11.1
  19. vLLM is vulnerable to Server-Side Request Forgery (SSRF) through `MediaConnector` class
    Wysoka · 7.1 CVE-2025-6242 >= 0.5.0, < 0.11.0 0.11.0
  20. vLLM is vulnerable to timing attack at bearer auth
    Wysoka · 7.5 CVE-2025-59425 < 0.11.0 0.11.0
  21. vLLM has remote code execution vulnerability in the tool call parser for Qwen3-Coder
    Wysoka · 8.8 CVE-2025-9141 >= 0.10.0, < 0.10.1.1 0.10.1.1
  22. vllm API endpoints vulnerable to Denial of Service Attacks
    Wysoka · 7.5 CVE-2025-48956 >= 0.1.0, < 0.10.1.1 0.10.1.1
  23. vLLM Allows Remote Code Execution via PyNcclPipe Communication Service
    Krytyczna · 9.8 CVE-2025-47277 >= 0.6.5, < 0.8.5 0.8.5
  24. Remote Code Execution Vulnerability in vLLM Multi-Node Cluster Configuration
    Wysoka · 8.0 CVE-2025-30165 >= 0.5.2, < 0.10.0 0.10.0
  25. vLLM Vulnerable to Remote Code Execution via Mooncake Integration
    Krytyczna · 10.0 CVE-2025-32444 >= 0.6.5, < 0.8.5 0.8.5
  26. Data exposure via ZeroMQ on multi-node vLLM deployment
    Wysoka · 7.5 CVE-2025-30202 >= 0.5.2, < 0.8.5 0.8.5
  27. CVE-2025-24357 Malicious model remote code execution fix bypass with PyTorch < 2.6.0
    Krytyczna · 9.8 ghsa-ggpf-24jw-3fcw < 0.8.0 0.8.0
  28. vLLM deserialization vulnerability in vllm.distributed.GroupCoordinator.recv_object
    Krytyczna · 9.8 CVE-2024-9052 <= 0.8.1 brak poprawki
  29. vLLM Deserialization of Untrusted Data vulnerability
    Krytyczna · 9.8 CVE-2024-11041 <= 0.6.2 brak poprawki
  30. vLLM allows Remote Code Execution by Pickle Deserialization via AsyncEngineRPCServer() RPC server entrypoints
    Krytyczna · 9.8 CVE-2024-9053 <= 0.6.0 brak poprawki
  31. vLLM Allows Remote Code Execution via Mooncake Integration
    Krytyczna · 9.0 CVE-2025-29783 >= 0.6.5, < 0.8.0 0.8.0
  32. vllm: Malicious model to RCE by torch.load in hf_model_weights_iterator
    Wysoka · 7.5 CVE-2025-24357 < 0.7.0 0.7.0
  33. vLLM denial of service vulnerability
    Wysoka · 7.5 CVE-2024-8768 < 0.5.5 0.5.5

Dowiedz się, które podatności dotyczą Twojej aplikacji

Po podłączeniu integracji DockRay i włączeniu raportowania oprogramowania DockRay porównuje zgłoszone wersje komponentów ze znanymi podatnościami i powiadamia o potwierdzonych dopasowaniach. Zakres zależy od obsługiwanej technologii i dostępnych biuletynów.

Podłącz swoją aplikację
Newsletter DockRay

Bezpieczeństwo aplikacji w Twojej skrzynce

Najważniejsze podatności, konkretne zalecenia i praktyczne wskazówki dla Twoich aplikacji. Wybierz technologie i śledź to, co ma znaczenie dla Twojego projektu.

  • Codzienne zestawienie nowych podatności, także krytycznych.
  • Powiadomienia o ważnych aktualizacjach.
  • Praktyczne porady o bezpieczeństwie i monitoringu aplikacji.

Pozostaw puste, aby otrzymywać informacje o wszystkich technologiach.

Polityka prywatności

Zapis jest bezpłatny. Wypiszesz się jednym kliknięciem w wiadomości.