CVE-2026-19889 – GitLab AI Gateway
“Crafted AI model routing can redirect trusted requests and expose cloud credentials and signing material.”
GitLab fixes two High-severity server-side request forgery vulnerabilities in the AI Gateway. CVE-2026-19889 allows an authenticated Duo Agent Platform user to redirect model requests to an attacker-controlled endpoint through crafted model metadata, potentially exposing Google Vertex AI or AWS Bedrock credentials. The CVSS score is 8.2, which is High severity.
CVE-2026-75871 allows a similar redirect through crafted inline flow configuration and Host header manipulation, potentially exposing Google Cloud Vertex credentials and private signing keys. The CVSS score is 8.2, which is High severity. Public proof-of-concept material is available for both vulnerabilities.
Key Details
- Attack Vector
- Network
- Attack Complexity
- High
- Privileges Required
- Low
- User Interaction
- None
- CWE Classification
- CWE-918