CVE-2025-46722: vLLM hash collisions in multimodal image hashing via raw pixel bytes

Summary

vLLM versions from 0.7.0 before 0.9.0 contain a flaw in the MultiModalHasher class in vllm/multimodal/hasher.py. Its image hashing method serializes PIL.Image.Image objects using only obj.tobytes(), which omits metadata such as width, height and mode, so images of different dimensions with identical pixel bytes can produce the same hash. This can cause hash collisions, incorrect cache hits, and possible data leakage.

Mitigation

Fixed in 0.9.0.

Affected packages in the Exposure Registry

Matched by package name and ecosystem. Each entry shows whether the package delegates to a language model and how many tracked packages depend on it.

• vllmPyPILLM dependency since 2023-08-25 · 15 tracked dependents

Topics

• Inference infrastructure

Related items

• MediumCVE-2026-121936: easyVoice server-side request forgery through openaiBaseUrl in /api/v1/ttsSimilar attack · NVD/CVE Database
• MediumCVE-2026-118628: MLflow get-artifacts server-side request forgery information disclosureSimilar attack · NVD/CVE Database
• HighCVE-2026-118625: MLflow Tracking Server privilege escalation via get_endpointsSimilar attack · NVD/CVE Database
• InfoWhen trust becomes the payload: Inside the threats, etc quarterly reportSimilar attack · Elastic Security Labs
• MediumCritical Langflow flaws enable authentication bypass and RCESimilar attack · CSO Online