Executive brief
Network-AI, a tool used for managing automated network actions, contains a security flaw where sensitive approval requests are exposed to unauthorized users. Even if a password is set, an attacker can view pending system commands, file paths, and justifications for high-risk actions. Additionally, a configuration error allows malicious websites to steal this information if an administrator visits them while logged into the tool.
Technical details
The vulnerability exists in the ApprovalInbox component of the network-ai npm package. While a previous fix (GHSA-mxjx-28vx-xjjj) implemented authorization for mutating POST requests, it failed to apply the 'checkAuth' logic to GET routes including /approvals/?status=all, /approvals/:id, /approvals/stats, and /approvals/sse. An unauthenticated attacker can access these endpoints to retrieve full ApprovalEntry details, such as shell-command strings and file paths. Furthermore, the application uses a hardcoded 'Access-Control-Allow-Origin: *' header, facilitating cross-origin data exfiltration via a victim's browser. The issue is patched in version 5.13.4.
Affected products
- Jovancoding Network-AI (npm: network-ai) 5.12.2 through 5.13.3
Timeline
- 2026-07-05: advisory: Vendor advisory GHSA-m4jg-6w3q-gm86 published
- 2026-07-20: disclosed: CVE-2026-64622 published to NVD
- 2026-07-20: patched: Version 5.13.4 released to address the incomplete fix