CVE-2024-10902 is a critical flaw (CVSS 9.8) in eosphoros-ai DB-GPT v0.6.0. The web API POST /v1/personal/agent/upload allows arbitrary file upload with path traversal, so files can land anywhere on the file system, potentially leading to remote code execution. No fix is confirmed, so restrict access.
What's affected
The CVE record lists eosphoros-ai/db-gpt version v0.6.0. The description says the agent upload API is vulnerable to arbitrary file upload with path traversal, allowing unauthorised attackers to upload files to the file system at any location. It names a malicious init.py in the Python site-packages directory as an example route to code execution. The reference is a huntr report that I could not read.
I could not confirm a fixed version from the sources available, so check the vendor site and release notes for a newer release. The sibling records are CVE-2024-10835 and CVE-2024-10901.
How to check if you're exposed
Read the version from your container image tag or project files. v0.6.0 is the version named in the record.
Then check exposure. A request to the upload endpoint from a client with no credentials shows whether your deployment requires a login. Do this only on a system you own, and use a harmless test file name in a scratch directory.
Check which directories the service account can write to. The wider those rights, the worse a path traversal upload becomes. Look through writable directories for files you did not put there, especially Python files.
What to do now
- Find DB-GPT instances and note their versions.
- Restrict the service to trusted networks, and add authentication in front of it.
- Run the service as a dedicated unprivileged user, with write access only to the directory where uploads belong.
- Check the project for a release newer than v0.6.0 and apply it.
- Audit writable directories for unexpected files, and rebuild the container from a clean image if you find any.
Why it matters
Path traversal (CWE-22) on an upload means the file name or path from the request decides where the file is saved, without being kept inside the intended folder. Combined with no access control, anyone who can reach the endpoint chooses where on the disk their file goes.
The three DB-GPT records show a pattern for AI tooling. Applications built quickly around language models often expose powerful internal APIs, such as running queries or saving files, and assume that only the intended user will call them. A public network breaks that assumption. The CVE is not on the CISA Known Exploited Vulnerabilities list, and I found no report of attacks in the wild.
Quick check
Why is the file location so important?
A file written into an application or package directory can be loaded and run later. A file in an isolated upload folder cannot.
Do I need to fix all three DB-GPT CVEs separately?
They affect different endpoints in the same version. A release that addresses them all is the goal, so check the project notes.
Sources
Last reviewed: 19 October 2026.
