Attack Chain

1. An unauthenticated remote attacker sends an HTTP POST request to a vulnerable h2oGPT OpenAI-compatible files API endpoint, such as /files/upload, /files/content, or /files/delete.
2. The attacker crafts the HTTP Authorization header to include a bearer token containing path traversal sequences (e.g., Bearer ../../../path/to/malicious_file.sh) to manipulate the file path.
3. Due to the default "EMPTY" API key, the h2oGPT server bypasses authentication checks and proceeds to process the request as if it were legitimate.
4. The get_user_dir function in openai_server/backend_utils.py is invoked, and it incorrectly incorporates the unsanitized bearer token string into a file path using os.path.join.
5. This misconfiguration allows the attacker to specify an arbitrary file path outside the intended user directory, leading to the ability to read, write, or delete files anywhere on the server where the h2oGPT process has write permissions.
6. The attacker uploads or modifies a malicious file (e.g., a web shell, a configuration file with injected commands, or a startup script) to a critical system location.
7. Upon the h2oGPT application restarting, reloading components, or when the tampered file is otherwise executed or loaded, the attacker's injected code achieves remote code execution (RCE) on the server.

Impact

Successful exploitation of CVE-2026-65700 grants unauthenticated remote attackers arbitrary file read, write, and delete capabilities on the compromised h2oGPT server. This directly leads to remote code execution (RCE) by allowing attackers to inject malicious code into application-loaded files or startup hooks. The consequences include complete compromise of the affected server, potential exfiltration of sensitive data, establishment of persistent backdoors, and lateral movement within the victim's network. Given the unauthenticated nature and the high CVSS score of 9.8, the vulnerability poses a critical threat to the confidentiality, integrity, and availability of systems running vulnerable h2oGPT versions.

Recommendation

• Patch CVE-2026-65700 immediately by upgrading h2oGPT to a version beyond 0.2.1 to mitigate this critical vulnerability.
• Deploy the Sigma rule Detect CVE-2026-65700 Exploitation Attempt - Path Traversal in Bearer Token to your web server logging infrastructure (e.g., Apache, Nginx access logs).
• Monitor web server access logs for HTTP POST requests to OpenAI-compatible API endpoints that contain path traversal sequences (../, ..\) within the cs-auth-token field or HTTP Authorization header.
• Ensure that all API keys, including default configurations, are replaced with strong, unique, and non-guessable credentials, and that API key authentication is properly enforced.