Executive brief
Network-AI is a tool used to coordinate multiple AI agents. A security flaw in its sandbox mechanism allows an agent or user to access files outside of their designated folder if the target folder shares a similar name prefix. This could lead to the unauthorized viewing or listing of sensitive system files that were intended to be protected by the sandbox boundary.
Technical details
A path traversal vulnerability exists in Network-AI's AgentRuntime and SandboxPolicy components due to improper path validation. The software uses a raw string prefix check (startsWith) to enforce sandbox boundaries, which fails to account for directory separators. An attacker can bypass the sandbox by targeting sibling directories that share the same name prefix as the configured basePath (e.g., accessing '/tmp/sandbox_evil' when the sandbox is set to '/tmp/sandbox'). This allows local users or agents to call readFile() or listDir() to access files outside the intended scope. The issue is resolved in version 5.12.2 by implementing separator-anchored prefix checks.
Affected products
- Jovancoding Network-AI < 5.12.2
Timeline
- 2026-06-18: patched: Fixed in version 5.12.2 via commit a59c13a
- 2026-07-20: disclosed: CVE-2026-58481 published