AI-Assisted Secure Coding (Appendix C) · Autonomous Agent Change Control Constraints
Separate people or systems at each stage
Verify that separation of duties holds across the stages of an AI-generated change. Each stage (generation, review, approval, deployment) is performed by a distinct principal, whether human or system.
OWASP AISVS v1.01, requirement AC.8.4
In plain English
An AI-generated change passes through four stages: generation, review, approval and deployment. A different principal, human or system, should perform each one. Then no single compromised identity can take a change from idea to production. This is a Level 3 control.
What good looks like
- Each of the four stages is carried out by a distinct identity.
- Tooling enforces the separation and records who did what.
- A single identity that appears in two stages of one change is blocked or flagged.
- Exceptions are recorded and approved.
How teams usually do it
- Separate identities and credentials per stage, with a pipeline check that compares the identities recorded in the change's provenance.
- Deployment environments with required reviewers who are neither author nor approver.
- The same identity rules applied to AI reviewers: a different system from the one that generated the code.
Suggested evidence
- The pipeline design and identity mapping.
- Provenance for one change showing four distinct principals.
Common mistakes
- One service account that generates and deploys.
- The same model instance, under two names, acting as author and reviewer.
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.