Sign in
Deutsch English Español Polski
Low

vLLM uses Python 3.12 built-in hash() which leads to predictable hash collisions in prefix cache

  • Disclosed: 06.02.2025
  • Updated: 11.10.2026 14:19

Summary

Maliciously constructed prompts can lead to hash collisions, resulting in prefix cache reuse, which can interfere with subsequent responses and cause unintended behavior.

Details

vLLM's prefix caching makes use of Python's built-in hash() function. As of Python 3.12, the behavior of hash(None) has changed to be a predictable constant value. This makes it more feasible that someone could try exploit hash collisions.

Impact

The impact of a collision would be using cache that was generated using different content. Given knowledge of prompts in use and predictable hashing behavior, someone could intentionally populate the cache using a prompt known to collide with another prompt in use.

Solution

We address this problem by initializing hashes in vllm with a value that is no longer constant and predictable. It will be different each time vllm runs. This restores behavior we got in Python versions prior to 3.12.

Using a hashing algorithm that is less prone to collision (like sha256, for example) would be the best way to avoid the possibility of a collision. However, it would have an impact to both performance and memory footprint. Hash collisions may still occur, though they are no longer straight forward to predict.

To give an idea of the likelihood of a collision, for randomly generated hash values (assuming the hash generation built into Python is uniformly distributed), with a cache capacity of 50,000 messages and an average prompt length of 300, a collision will occur on average once every 1 trillion requests.

References

Affected software

Click a technology to see all vulnerabilities affecting it.

PlatformPackageVulnerable versionsFixed in
Python (PyPI) vllm < 0.7.2 0.7.2

References

CVSS score 2.6/10
CVSS vector
CVSS:3.1/AV:N/AC:H/PR:L/UI:R/S:U/C:N/I:L/A:N
Weakness (CWE)
CWE-354
Platform
Python (PyPI)
Package
vllm
Disclosed
06.02.2025
Updated
11.10.2026 14:19

Sources and licenses

OSV / GitHub Advisory Database - PYSEC-2025-62 - PyPA Advisory Database; CC BY 4.0

OSV / GitHub Advisory Database - GHSA-rm76-4mrf-v9r8 - CC BY 4.0

Vulnerability history: vllm

All published vulnerabilities of this package (Python (PyPI)), newest first. Shows whether this is a one-off slip or a package that needs watching.

All (94) in the archive
Total entries
94
First → latest
09.2024 10.2026
Without a fix
5
Exploited in attacks
0

Severity breakdown

  • Critical 9
  • High 23
  • Medium 47
  • Low 4
  • Unrated 11

Disclosures per year

Severity:
  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
    High · 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
    High · 7.5 CVE-2026-100652 >= 0.22.0, < 0.24.0 0.24.0
  3. vLLM has Remote DoS via Invalid Recovered Token Reinjection
    High · 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
    High · 7.5 CVE-2026-55574 < 0.24.0 0.24.0
  5. vLLM denial of service via prompt embeds on M-RoPE models
    High 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.
    High · 8.8 CVE-2026-54232 < 0.22.1 0.22.1
  7. vLLM: OpenAI auth bypass
    Critical · 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
    High · 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
    High · 7.5 CVE-2026-5497 >= 0.8.0, < 0.19.0 0.19.0
  10. vLLM has Hardcoded Trust Override in Model Files Enables RCE Despite Explicit User Opt-Out
    High · 8.8 CVE-2026-27893 >= 0.10.1, < 0.18.0 0.18.0
  11. vLLM has RCE In Video Processing
    Critical · 9.8 CVE-2026-22778 >= 0.8.3, < 0.14.1 0.14.1
  12. vLLM vulnerable to Server-Side Request Forgery (SSRF) through MediaConnector
    High · 7.1 CVE-2026-24779 < 0.14.1 0.14.1
  13. vLLM affected by RCE via auto_map dynamic module loading during model initialization
    High · 8.8 CVE-2026-22807 >= 0.10.1, < 0.14.0 0.14.0
  14. vLLM introduced enhanced protection for CVE-2025-62164
    High · 8.8 CVE-2026-56340 >= 0.10.2, < 0.13.0 0.13.0
  15. vLLM vulnerable to remote code execution via transformers_utils/get_config
    High · 7.1 CVE-2025-66448 < 0.11.1 0.11.1
  16. vLLM vulnerable to DoS with incorrect shape of multimodal embedding inputs
    High · 6.5 CVE-2025-62372 >= 0.5.5, < 0.11.1 0.11.1
  17. vLLM deserialization vulnerability leading to DoS and potential RCE
    High · 8.8 CVE-2025-62164 >= 0.10.2, < 0.11.1 0.11.1
  18. vLLM is vulnerable to Server-Side Request Forgery (SSRF) through `MediaConnector` class
    High · 7.1 CVE-2025-6242 >= 0.5.0, < 0.11.0 0.11.0
  19. vLLM is vulnerable to timing attack at bearer auth
    High · 7.5 CVE-2025-59425 < 0.11.0 0.11.0
  20. vLLM has remote code execution vulnerability in the tool call parser for Qwen3-Coder
    High · 8.8 CVE-2025-9141 >= 0.10.0, < 0.10.1.1 0.10.1.1
  21. vllm API endpoints vulnerable to Denial of Service Attacks
    High · 7.5 CVE-2025-48956 >= 0.1.0, < 0.10.1.1 0.10.1.1
  22. vLLM Allows Remote Code Execution via PyNcclPipe Communication Service
    Critical · 9.8 CVE-2025-47277 >= 0.6.5, < 0.8.5 0.8.5
  23. Remote Code Execution Vulnerability in vLLM Multi-Node Cluster Configuration
    High · 8.0 CVE-2025-30165 >= 0.5.2, < 0.10.0 0.10.0
  24. vLLM Vulnerable to Remote Code Execution via Mooncake Integration
    Critical · 10.0 CVE-2025-32444 >= 0.6.5, < 0.8.5 0.8.5
  25. Data exposure via ZeroMQ on multi-node vLLM deployment
    High · 7.5 CVE-2025-30202 >= 0.5.2, < 0.8.5 0.8.5
  26. CVE-2025-24357 Malicious model remote code execution fix bypass with PyTorch < 2.6.0
    Critical · 9.8 ghsa-ggpf-24jw-3fcw < 0.8.0 0.8.0
  27. vLLM deserialization vulnerability in vllm.distributed.GroupCoordinator.recv_object
    Critical · 9.8 CVE-2024-9052 <= 0.8.1 no fix
  28. vLLM Deserialization of Untrusted Data vulnerability
    Critical · 9.8 CVE-2024-11041 <= 0.6.2 no fix
  29. vLLM allows Remote Code Execution by Pickle Deserialization via AsyncEngineRPCServer() RPC server entrypoints
    Critical · 9.8 CVE-2024-9053 <= 0.6.0 no fix
  30. vLLM Allows Remote Code Execution via Mooncake Integration
    Critical · 9.0 CVE-2025-29783 >= 0.6.5, < 0.8.0 0.8.0
  31. vLLM uses Python 3.12 built-in hash() which leads to predictable hash collisions in prefix cache this entry
    Low · 2.6 CVE-2025-25183 < 0.7.2 0.7.2
  32. vllm: Malicious model to RCE by torch.load in hf_model_weights_iterator
    High · 7.5 CVE-2025-24357 < 0.7.0 0.7.0
  33. vLLM denial of service vulnerability
    High · 7.5 CVE-2024-8768 < 0.5.5 0.5.5

Know which vulnerabilities affect your application

With the DockRay integration and software inventory enabled, DockRay checks reported component versions against known vulnerabilities and notifies you about confirmed matches. Coverage depends on the supported technology and available advisory data.

Connect your application
DockRay newsletter

Application security in your inbox

Important vulnerabilities, actionable guidance and practical advice for your applications. Choose your technologies and follow what matters to your project.

  • A daily digest of new vulnerabilities, critical ones included.
  • Notifications about important updates.
  • Practical advice on application security and monitoring.

Leave empty to receive updates for every technology.

Privacy policy

Free to subscribe. Unsubscribe with one click in any email.