CognicellAI's Cognition platform, before version 0.13.0, fails to require authentication on its REST and SSE interfaces and defaults to a local execution sandbox. This allows an unauthenticated user to create a session and send commands that the built‑in Deep Agents ReAct loop executes in a shell on the host, delivering code that runs with the server process privileges. The result is uncontrolled remote code execution with full access to the server's operating system, making the vulnerability available to malicious actors remotely without any credentials.

The affected product is CognicellAI's Cognition, all releases earlier than 0.13.0. The CVE applies to configurations where the COGNITION_ALLOW_UNSAFE_LOCAL_EXECUTION flag is not explicitly set to true, and the application is exposed over a network (HTTP/HTTPS). No specific sub‑versions are listed beyond the 'less than 0.13.0' boundary.

The vulnerability carries a CVSS base score of 9.2, indicating critical severity. The EPSS score is not reported, and the issue is not currently listed in the CISA KEV catalog. Attackers can exploit the flaw over standard HTTP endpoints (POST /sessions and POST /sessions/{id}/messages) from any remote network, as no authentication is required. Exploitation would provide arbitrary code execution on the system hosting the Cognition service, with the same privileges as the server process.