Network‑AI versions older than 5.15.1 allow attackers to craft command strings that use quotes to hide malicious payloads from the SandboxPolicy blocklist but are executed unquoted by the underlying executor. This mismatch enables the bypass of approval gates and can result in the execution of arbitrary commands with the privileges of the running process, exposing confidentiality, integrity, and availability of the system. The weakness corresponds to CWE‑436, representing a mismatch in semantic interpretation of inputs across subsystems.

Jovancoding Network‑AI, all deployments with a build prior to 5.15.1 are affected. No specific sub‑versions or additional product families were identified in the advisory.

The evaluated CVSS score of 8.7 indicates high severity and the absence of an EPSS rating suggests no publicly known exploit patterns, but the vulnerability is listed as not in the CISA KEV catalog. The attack vector is inferred to be remote, as the blocklist protection typically intercepts inputs received through network interfaces or uploaded files. Exploitation requires an attacker to supply a crafted command string that passes the blocklist check but is re‑tokenized without quotes for execution, allowing the injection of dangerous arguments. Given the high CVSS, the risk of compromise is significant if the system is exposed. Continued monitoring for new exploits is advised.