FastGPT flaws enable GitHub Actions code execution and NoSQL auth bypass

FastGPT, an AI agent building platform, was disclosed with multiple high-severity vulnerabilities affecting both its application and development pipeline. CVE-2026-33075 impacts versions 4.14.8.3 and earlier and stems from the fastgpt-preview-image.yml GitHub Actions workflow using pull_request_target while checking out untrusted fork code. The flaw allows external contributors to achieve arbitrary code execution in GitHub Actions, exfiltrate secrets, and potentially trigger a supply-chain compromise by building and pushing attacker-controlled container images to the production registry. At disclosure, no patch was available for that issue; GitHub tracked it as GHSA-xfx8-w35j-485c.

Two additional flaws, CVE-2026-40351 and CVE-2026-40352, affect FastGPT versions prior to 4.14.9.5 and were fixed in that release. Both are NoSQL injection bugs caused by missing runtime validation in password-handling logic. The first lets an unauthenticated attacker bypass login checks by supplying MongoDB operators such as a password object matching any value, enabling login as arbitrary users including the root administrator. The second lets an authenticated low-privileged user bypass old-password verification in the password-change endpoint, enabling unauthorized password resets and possible account takeover, with broader impact if combined with ID manipulation.