CVE-2026-54051 is an OS command injection vulnerability in Jovancoding's Network-AI prior to version 5.9.1. The agent sandbox uses an allowlist that glob-matches entire command strings, but executes commands through '/bin/sh -c'. This allows wildcard patterns like 'git *' or 'npm *' to match commands with appended shell metacharacters (e.g., 'git status; <anything>'), enabling arbitrary command execution. The vulnerability is addressed in version 5.9.1 by switching to spawn execution without a shell and by rejecting unquoted shell metacharacters before allowlist matching. The fix prevents shell injection by parsing arguments properly and enforcing stricter command validation.

An attacker with limited privileges can exploit the allowlist wildcard matching combined with shell execution to run arbitrary OS commands, leading to full compromise of confidentiality, integrity, and availability of the affected system. The CVSS score of 9.9 reflects critical impact with network attack vector, low complexity, and no user interaction required.

Upgrade Network-AI to version 5.9.1 or later, where the vulnerability is fixed by eliminating shell execution and enforcing strict command tokenization. Additionally, avoid using broad wildcard allowlist entries such as 'node *' or 'npm *' as they inherently allow arbitrary code execution by design. No other vendor advisories or patches are indicated; users should rely on the official upgrade.