Based on Appendix C, “AI-Assisted Secure Coding,” of OWASP AISVS, this article explains 5 controls for the secure use of AI coding tools and the key points of each.
As coding tools powered by generative AI become more widespread, security risk has emerged as a challenge. AISVS covers security validation for AI systems as a whole, while Appendix C organizes 14 practical controls for development environments.
This blog focuses on 5 of these controls: AC.3, “Protection of Prompts and Context”; AC.4, “Validation of AI-Generated Code”; AC.8, “Change Management Constraints for Autonomous AI Agents”; AC.11, “Hardening AI Code Review Bots”; and AC.12, “Hardening AI-Aware CI/CD Pipelines.” It organizes the key points from perspectives including the handling of external content, pre-merge validation and ensuring the security of automation.
AC.3 calls for preventing the inclusion of confidential information and screening external inputs. AC.4 calls for an approval workflow combining human review with automated scanning. AC.8 requires thorough separation of duties to technically prevent self-approval by autonomous agents. AC.11 calls for treating repository inputs as untrusted, protecting bot settings and managing permissions. AC.12 emphasizes protecting secrets in CI/CD pipelines and making security reviews mandatory.
These 5 controls are positioned as specific measures for mitigating new risks arising from the adoption of AI and are considered useful for understanding the practical application of AISVS.