AI-Assisted Secure Coding (Appendix C) · AI-Generated Infrastructure & Pipeline Artifacts
Human approval before AI infrastructure runs
Verify that AI-generated infrastructure and pipeline configurations require human review and approval before they run in any environment beyond a hermetic sandbox.
OWASP AISVS v1.01, requirement AC.7.2
In plain English
AI-generated infrastructure and pipeline changes need a person to review and approve them before they run anywhere beyond a sealed sandbox. An agent that can write and apply infrastructure code can open a firewall or delete a database before anyone has read the change.
What good looks like
- Apply and deploy steps for AI-generated changes wait for human approval.
- Only an isolated sandbox, with no route to real data or shared systems, runs them unapproved.
- The reviewer sees the real plan or diff.
- Approvals are logged.
How teams usually do it
- A plan-then-apply workflow: the pipeline produces a plan, a person reviews it, and only then does apply run. Terraform Cloud, Atlantis and deployment environments with required reviewers all support this.
- Agent credentials that allow plan and deny apply.
- Disposable sandbox accounts for agent experiments.
Suggested evidence
- Pipeline configuration.
- Approval logs.
- The agent identity's permissions.
Common mistakes
- Auto-applying AI-generated Terraform in a shared development account.
- A "sandbox" that shares networks or credentials with real environments.
- Approving the PR description without reading the plan.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.