Zero Day Initiative Discloses LiteLLM Command Injection Vulnerability
A critical command injection vulnerability in LiteLLM, assigned ZDI-26-1031, allows authenticated attackers to execute arbitrary code remotely.

The Zero Day Initiative (ZDI) has disclosed a critical command injection vulnerability, tracked as ZDI-26-1031, affecting LiteLLM. This flaw permits authenticated remote attackers to execute arbitrary code on vulnerable installations.
The vulnerability stems from improper validation of user-supplied strings within the handling of MCP stdio configurations. When these strings are used in system calls without adequate sanitization, an attacker can inject malicious commands, leading to code execution in the context of the service account.
Exploitation of this vulnerability requires prior authentication to the LiteLLM instance. While this adds a barrier, the potential impact of arbitrary code execution remains severe, allowing attackers to compromise the underlying system.
ZDI initially reported the vulnerability to the vendor on June 17, 2026. The disclosure timeline reveals a protracted engagement, with the vendor initially closing the report as a false positive on September 15, 2026. ZDI disputed this closure, highlighting an incomplete fix, which led the vendor to reopen the report. Despite vendor assurances of a planned fix and GHSA before the November 18th publication deadline, the vendor ultimately clarified that no separate patch would be released and disputed the existence of a vulnerability.
In response to the vendor's stance, ZDI proceeded with the publication of a zero-day advisory on November 18, 2026. The vendor's recommended mitigation involves upgrading to version 1.104.0 or later and ensuring that the LITELLM_ENABLE_MCP_STDIO environment variable is not set, as this feature is disabled by default in newer versions unless explicitly enabled. Additionally, users are advised to replace default master keys with strong, random values, a measure enforced in v1.104.0+.
The vulnerability was discovered by David Fiser and Alfredo Oliveira of TrendAI Research. The CVSS score assigned by ZDI is 7.2, indicating a high severity.
This disclosure highlights the ongoing challenges in vulnerability management and vendor-product interactions, particularly when dealing with complex software stacks like those used in AI model serving. The lack of a clear patch and the vendor's initial dispute underscore the importance of independent security research and disclosure processes.
Users of LiteLLM are strongly urged to review the vendor's mitigation advice, upgrade to the latest version, and secure their configurations to prevent potential exploitation of this critical command injection flaw.