Checklist

LAYER 1: STANDARDS AND GROUNDING

Set up the foundation of your AI system by providing documentation of your coding and security standards.

• Document your organization’s coding standards and provide them to the AI before generating any code.
• Give your organization’s security standards and selected OWASP controls to the AI.
• Provide architecture decision records and security tradeoff documentation to the AI at the start of every session.
• Communicate regulatory and compliance requirements (PCI-DSS, HIPAA, SOC 2, etc.) to the AI at the start of every AI session that involves sensitive data flows.
• Use precise, security-explicit prompts that instruct the AI to optimize for safety, not just correctness.

LAYER 2: THREAT MODELING

Develop separate threat models for build time and runtime to limit insecure code generation and AI-specific risks.

• Create a build time threat model for AI-introduced vulnerabilities like insecure code generation, hallucinated packages, and prompt injection.
• Build a runtime threat model covering prompt injections, data leakage, hallucinations, and model supply chain risks.
• Model these agentic AI threat scenarios explicitly: autonomous commits, infrastructure changes, pipeline triggers, and permission grants without human authorization.

LAYER 3: AI-GENERATED TESTS AND SECURITY LOOPS

Create a security loop that requires the AI to run tests at every stage of development. Never treat AI-generated code as final.

• Require the AI to generate unit tests for all AI-generated code before it is considered complete.
• Include regression tests for every AI-generated bug fix to prevent the same vulnerability from re-entering the codebase.
• Require integration tests covering AI component boundaries, database interactions, external service calls, and trust boundary crossings.
• Separate bug reviews from security reviews. They require different prompts.
• Run tests at every stage of the AI coding cycle, not just before release.
• Verify that AI-generated tests exercise meaningful security properties, not just surface-level pass conditions.

LAYER 4: AI REVIEWS AI

Add a layer of scrutiny to every change before it reaches human reviewers by creating a separate AI pass specifically scoped to security.

• Deploy a dedicated AI security review for only changed code, not the entire codebase.
• Configure AI security reviews to run automatically on every pull request before human review begins.
• Make sure AI review output feeds into your developer workflow as actionable, annotated findings, not a separate security report that gets ignored.
• Track AI-flagged security findings over time and use the data to improve the grounding and standards documents provided to the system in Layer 1.

LAYER 5: SAST, SCA, AND DAST

Utilize the three main scanning technologies to ensure that remediation happens before any public exposure.

• Run SAST on every commit in a CI gate, blocking merges on any critical findings.
• Deploy IDE-based SAST scanning to catch insecure AI-generated code at the point of generation and before it enters the pipeline.
• Run SCA on every commit to detect hallucinated, nonexistent, and maliciously registered packages before they enter the build.
• Verify that every AI-suggested package exists in an approved registry before installation or before it is added to the dependency manifest.
• Run DAST against the running application to catch the gap between intent and behavior that even thorough code reviews can miss.
• Run SAST, SCA, and DAST on every commit if the pipeline can support it, and always before any major release.
• Ensure that AppSec tooling provides unified, cohesive findings across SAST, SCA, and DAST scans rather than siloed, disconnected outputs.

LAYER 6: HUMAN SECURITY REVIEWS

Have human developers and security experts use their valuable judgment to perform mandatory reviews.

• Require a security engineer or senior developer to review every AI-generated change for payment flows, authentication logic, and authorization boundaries.
• Require human review with explicit trust boundary analysis for all agentic AI workflows before granting them production access or real-world permissions.
• Train security reviewers to apply heightened scrutiny to AI-generated code by challenging assumptions, independently verifying outputs, and carefully assessing whether sensitive data is being handled appropriately.
• Maintain end-to-end visibility for development, DevOps, and AppSec teams across every stage of the SDLC, from IDE through runtime.

7: RUNTIME DEFENSES

Defend runtime AI systems against attacks that don’t occur in traditional deployments.

• Deploy runtime monitoring and anomaly detection specifically tuned to AI-generated application behaviors and agentic system actions.
• Check for prompt injections at inference time for any AI system that processes external, user-controlled, or third-party content.
• Monitor model supply chain integrity: Track model versions, providers, and API endpoints used by AI-enabled systems in production.
• Implement human-in-the-loop gates for all agentic AI actions with real-world consequences.
• Apply least-privilege principles to agentic AI systems: Grant the minimum permissions required and restrict access to production databases, secrets, and infrastructure by default.
• Maintain incident response runbooks that cover AI-related security events including prompt injection exploitation, model misbehavior, and hallucinated package compromise.

KEEPING IT ALL TOGETHER: CONTINUOUS GOVERNANCE

If these seven layers represent a cake, consider governance the frosting that holds everything together.

Maintain security by having a robust governance strategy that establishes AI usage policies, guidelines, and standards.

• Create and publish an organizational policy for approved AI coding assistants, usage boundaries, and data-handling requirements.
• Implement monitoring and detection of unauthorized AI tool usage within development environments.
• Establish a clear process for developers to request approval for new AI tools to reduce shadow IT.
• Formally assess and document your organization’s current readiness to secure AI-generated code.
• Actively build a culture that treats AI output as a starting point requiring verification instead of a finished product ready for deployment.
• Recognize and reward developers who find and call out AI-generated security issues.
• Evolve DevSecOps practices in parallel with AI adoption to ensure that security is embedded at every stage of the AI-accelerated SDLC.