Crucially, an authentication bypass in dbgpt_serve/utils/auth.py allows unauthenticated users to access this API endpoint. By submitting a crafted filename containing parent directory references (e.g., ../../../) or absolute paths, an attacker can write arbitrary files to the server's filesystem. An attacker can leverage this primitive to drop a malicious Python module into an application directory. Given the application's structure, the server process will execute the attacker-controlled code upon the subsequent import of the corrupted module. This vulnerability allows for full system compromise from an unauthenticated remote position.

Attack Chain

1. Attacker identifies the target DB-GPT instance reachable via the network.
2. Attacker crafts a multipart HTTP request targeting the skill_upload API endpoint.
3. Attacker sets the multipart filename attribute to a path-traversal string, such as ../../../path/to/malicious_module.py.
4. Attacker includes the payload (Python code) in the request body to be written to the target location.
5. The server process, lacking filename validation, writes the file to the malicious destination on the filesystem.
6. The application performs a subsequent operation or import that loads the newly created .py file.
7. The Python interpreter executes the attacker-controlled script within the context of the server process.
8. Attacker gains persistent remote code execution and potential full system compromise.

Impact

Successful exploitation of CVE-2026-80104 leads to unauthenticated Remote Code Execution (RCE) on the DB-GPT server. This allows an attacker to execute arbitrary system commands, steal sensitive data, or install persistent backdoors. Given that DB-GPT often handles AI/ML configurations and model artifacts, the impact includes potential exfiltration of proprietary datasets and credentials.

Recommendation

• Immediately restrict network access to the DB-GPT API endpoint to trusted internal networks only.
• Update DB-GPT to the patched version that implements filename canonicalization and validates that the destination path resides within the intended upload directory.
• Audit the packages/dbgpt-app/src/dbgpt_app/openapi/api_v1/agentic_data_api.py module for improper path concatenation practices.
• Deploy webserver logs monitoring to detect POST requests to the skill_upload endpoint that contain directory traversal patterns (e.g., ..%2f or ../).
• Implement stricter authentication checks in dbgpt_serve/utils/auth.py to ensure only verified users can access administrative API endpoints.