QuickGuard

Quick guard -A security rehearsal and release platform for AI infrastructure. It uses GitLab Duo agents to investigate incidents, reproduce them in isolated environments, propose repairs, and test security policies before deployment. The problem it solves:- Consider an enterprise running an AI agent that retrieves documents, executes Python, calls tools, and accesses an inference server. A security engineer introduces an outbound network restriction to prevent data leakage. That change might also interrupt model downloads, package installation, or permitted API calls. GuardStage tests the proposed restriction against both prohibited and legitimate operations. GitLab Duo then uses the results to prepare a more precise repair, with evidence attached to the merge request.

How the product works

1. Receive an incident or proposed change. Start from a GitLab issue, policy merge request, failed security test, or incident imported through an integration.
2. Build a reproducible scenario. Collect the relevant application version, policy, workload identity, dependency versions, and sanitized inputs. Record missing context explicitly.
3. Create isolated comparison environments. The orchestrator provisions a baseline environment and a candidate environment. Each receives the same fixtures and workload sequence.
4. Evaluate security and functionality. Check expected denials, permitted operations, application results, resource usage, and latency.
5. Use Duo to prepare a repair. Agents inspect failures, propose changes, and open an MR. The pipeline reruns the comparison against the revised candidate.
6. Release and verify. Package the passing configuration, deploy through the selected approval process, and monitor the actual outcome.

Built With

• claude
• gitlabduo
• google-cloud
• llm
• opa
• vllm