Remote Code Execution in Xinference API for Speech and Multimodal Models
CVE-2026-61539

10CRITICAL

Key Information:

Vendor

Xorbitsai

Status
Inference
Vendor
CVE Published:
21 August 2026

What is CVE-2026-61539?

The Xinference API, used for running open-source speech and multimodal models, contains a vulnerability wherein it passes untrusted output from the Llama3 tool calls to the eval() function. This security flaw allows unauthenticated remote attackers to craft specific prompts that influence the model's output, leading to the unintended execution of commands within the Xinference server environment. The issue has been addressed in version 2.7.0, patching the path through which these commands could be executed, ensuring a more secure operation of the API.

Affected Version(s)

inference < 2.7.0

References

https://github.com/xorbitsai/inference/security/advisorie...
x_refsource_CONFIRM
https://github.com/xorbitsai/inference/pull/4786
x_refsource_MISC
https://github.com/xorbitsai/inference/commit/1b3d220f342...
x_refsource_MISC

CVSS V3.1

Score:
10
Severity:
CRITICAL
Confidentiality:
High
Integrity:
High
Availability:
High
Attack Vector:
Network
Attack Complexity:
Low
Privileges Required:
None
User Interaction:
None
Scope:
Changed

Timeline

  • Vulnerability published

  • Vulnerability Reserved

.