Network-AI versions before 5.12.2 use raw string prefix tests to enforce sandbox directory boundaries in AgentRuntime's file access methods. This approach incorrectly allows access to sibling directories that share the sandbox base path prefix, enabling path traversal outside the intended restricted directory. The vulnerability affects calls to AgentRuntime.readFile() and AgentRuntime.listDir(), breaking the documented sandbox boundary. The fix in version 5.12.2 updates SandboxPolicy.resolvePath() and isPathAllowed() to use separator-anchored prefix checks, ensuring only the exact base directory or its true subdirectories are accessible, thus mitigating the path traversal issue.

An attacker or agent with permission to invoke AgentRuntime.readFile() or AgentRuntime.listDir() can bypass sandbox restrictions and access files outside the intended directory. This leads to unauthorized disclosure of potentially sensitive files. The vulnerability does not impact integrity or availability but has a high confidentiality impact. The CVSS 3.1 base score is 6.5 (medium severity), reflecting local attack vector with low complexity and no user interaction required.

Upgrade Network-AI to version 5.12.2 or later, where the path traversal vulnerability is fixed by improved path containment checks. Versions prior to 5.12.2 are vulnerable. No other mitigations are documented. Patch status is not explicitly stated but the fix is included in version 5.12.2, so applying this official update is the recommended remediation.