AI enriched 16 days ago (2026-04-09 06:44 UTC)

Technical Summary

vLLM is an inference and serving engine for large language models (LLMs). In versions from 0.6.4 to before 0.12.0, users can crash the vLLM engine serving multimodal models that use the Idefics3 vision model implementation by sending a specially crafted 1x1 pixel image. This causes a tensor dimension mismatch that results in an unhandled runtime error, leading to complete server termination. This issue has been patched in version 0.12.0.

Potential Impact

Severity is MEDIUM (CVSS 6.5). Depending on deployment context, affected components may be exposed to unauthorized actions or data integrity risk.

Exploitability Assessment

Exploitability is assessed as Medium based on published exploit references.

Primary risk drivers: published exploit references

Mitigation Recommendations

Validate affected product versions, prioritize patching, and monitor references for vendor remediation guidance. If immediate patching is not possible, apply compensating controls and limit exposure of vulnerable surfaces.

Detection & Monitoring

Track authentication anomalies, unexpected file writes, and suspicious plugin API activity around affected components.

Business Impact Lens

Prioritize remediation where affected components process customer data, admin sessions, or Internet-exposed workflows.