CVE-2026-105754 vllm-project CVE debrief

Who should care

Defenders responsible for large language model inference and serving engines, particularly those using vLLM, should assess exposure and prioritize updates to version 0.30.0 or later.

Why it matters

Defenders should care about CVE-2026-105754 as it can lead to denial of service or unintended behavior in vLLM, particularly in large language model inference and serving engines. The issue requires verification of input validation and multimodal field processor rebinding to the active model renderer contract. Updating to version 0.30.0 or later is recommended.

• Potential denial of service due to forged grid geometry or field types
• Possible poison or retrieval of cross-request encoder-cache state via forged cache hashes
• Altered replayed transport semantics due to dropped sparse placeholder masks
• Verification of input validation and multimodal field processor rebinding to the active model renderer contract

Technical summary

The /inference/v1/generate endpoint in the disaggregated scale-out path of vLLM accepts caller-supplied tensors, cache identifiers, and multimodal field processors without proper validation. This oversight allows for potential denial of service or unintended behavior due to forged grid geometry, field types, or non-positive placeholder lengths. The issue requires verification of input validation and multimodal field processor rebinding to the active model renderer contract. Updating to version 0.30.0 or later addresses the vulnerability.

Defensive priority

Defenders should prioritize updating to version 0.30.0 or later to address the vulnerability.

Recommended defensive actions

• Update to version 0.30.0 or later
• Review and validate input to the /inference/v1/generate endpoint
• Monitor for unusual activity or errors related to the vLLM engine
• Confirm whether affected product deployments exist in managed environments and assign an owner for follow-up
• Review compensating controls for exposed systems while remediation is scheduled and verified
• Check relevant monitoring, detection, and logs for exposed assets that need extra review
• Track exceptions, retest remediated assets, and close the item only after evidence is documented

Evidence notes

The CVE record and source item provide details on the vulnerability, including its description, affected versions, and fixed version. Evidence is limited to public sources and may not cover all affected or potentially affected deployments. Defenders should verify input validation practices, review multimodal field processor implementation, and monitor for unusual activity or errors related to the vLLM engine. Additional verification may be required based on specific deployment configurations and potential exposure.

Sources and references

Methodology and review provenance

AI-assisted synthesis based on stored public vulnerability evidence. System validation, approval state, and publication status do not by themselves establish human review of this revision. PatchSiren helps prioritize defensive review and does not prove exposure or remediation on any system.