DevOps software maintainers have distributed emergency security updates to resolve a maximum severity template sandbox escape vulnerability in self-hosted artificial intelligence services. Authenticated remote actors can execute arbitrary system commands on underlying gateway host servers.
The vulnerability, tracked as CVE-2026-90970 with a CVSS score of 9.9, impacts self-hosted instances of GitLab AI Gateway powering GitLab Duo. The defect resides within template rendering subroutines processing Duo Agent Platform queries. Authenticated users with access to agent workflows submit crafted template payloads to break sandbox isolation boundaries, executing operating system commands with elevated process privileges on the hosting server node. Official security bulletins urge administrators managing self-hosted AI Gateways to apply patched maintenance builds immediately.
Subverting artificial intelligence processing gateways introduces severe software supply chain exposure across enterprise development pipelines. Because AI Gateways integrate directly with source code repositories, continuous integration runners, and deployment secrets, an unauthenticated or low-privilege host takeover enables threat actors to extract repository secrets and poison production software builds.
– Force immediate maintenance upgrades across all self-hosted GitLab AI Gateway deployments to official fixed release builds.
– Gate public internet access to AI Gateway management endpoints using zero trust access control proxies.
– Inspect AI Gateway execution logs for unverified template evaluation calls or anomalous process spawns.
– Rotate service account tokens and API keys processed within continuous integration build pipelines.
DevOps platform defense demands rapid dependency patch execution and strict template sanitization to ensure self-hosted artificial intelligence gateways remain protected against arbitrary command execution. #CodeDefence #GitLab #DevSecOps #AISecurity #SupplyChain #RCE #AppSec #PatchManagement
/
