A security assessment of AI-generated codebases revealed a consistent pattern: each repository contained an average of 15 vulnerabilities. This finding emerged from analyzing production code written by AI coding assistants across multiple organizations. The surprise wasn't the number itself, but what influenced it most.
What Happened
Development teams using AI code generators to speed up feature delivery found their codebases riddled with systematic security flaws. These vulnerabilities weren't random; they clustered around specific framework choices, regardless of the AI model used. Teams using GitHub Copilot, Amazon CodeWhisperer, or other assistants saw similar vulnerability counts when paired with the same frameworks.
The pattern held across SQL injection risks, authentication bypasses, and insecure deserialization flaws. Framework selection determined vulnerability density more than the sophistication of the AI model.
Timeline
Initial Discovery Phase
Security teams noticed elevated vulnerability counts during routine application security testing. The flaws appeared in recently shipped features built with AI assistance.
Pattern Recognition
After reviewing multiple codebases, analysts identified that framework choice correlated more strongly with vulnerability presence than the AI tool itself. Teams using Framework A with any AI model saw similar flaw patterns to other teams using Framework A with different models.
Impact Assessment
Organizations running these codebases in production environments had already exposed the vulnerabilities to potential exploitation. Some flaws had existed for weeks before detection.
Which Controls Failed or Were Missing
No Pre-Commit Security Validation
Teams accepted AI-generated code directly into repositories without automated security checks. OWASP ASVS v4.0.3 Requirement 14.2.1 mandates that all code must be reviewed for security defects before integration. Most teams skipped this step for AI-generated code, treating it as equivalent to human-reviewed output.
Missing Framework Security Baselines
Organizations lacked documentation defining which frameworks produce secure-by-default configurations when paired with AI generators. ISO 27001 Control A.8.1 requires organizations to inventory and classify assets, including development frameworks. Without this inventory, teams couldn't assess framework-specific risk profiles.
Inadequate Code Review Process
Human reviewers focused on functionality, not security implications. PCI DSS v4.0.1 Requirement 6.3.2 requires security reviews of custom code prior to release. The reviews occurred, but reviewers didn't understand how AI-generated code patterns interacted with specific frameworks to create vulnerabilities.
No AI-Specific Secure Coding Standards
Teams applied the same secure coding guidelines to AI-generated and human-written code. NIST 800-53 Rev 5 Control SA-11 requires developer security testing and evaluation. The control assumes human developers who understand context. AI generators lack this context awareness, particularly around framework-specific security features.
What the Relevant Standard Requires
OWASP ASVS v4.0.3 Section 1.14: Configuration
Verify that the application framework is configured to use strong cryptography and that the default configuration is secure. AI generators often use framework defaults without verifying their security posture. When a framework's default session management is weak, the AI will implement it exactly as documented, creating vulnerabilities.
PCI DSS v4.0.1 Requirement 6.5.3
"Insecure cryptographic storage" must be prevented through secure coding practices. AI-generated code frequently implements cryptographic operations using framework convenience methods that don't meet this requirement. The AI prioritizes code that compiles and runs over code that satisfies security requirements.
ISO/IEC 27001:2022 Control A.8.28
"Secure coding" practices must be applied to software development. This includes understanding how frameworks handle security-sensitive operations. AI generators can't distinguish between a framework's secure API and its legacy compatibility API. They choose based on code pattern frequency in training data.
Lessons and Action Items for Your Team
Map Your Framework Risk Profile
Document which frameworks you're using and research their security track records with AI-generated code. Create a matrix: Framework X + AI Tool Y = Known Vulnerability Patterns. Update this quarterly as new patterns emerge.
Implement Pre-Commit Security Gates
Add static analysis tools that understand framework-specific vulnerabilities to your CI/CD pipeline. Configure them to flag AI-generated code for additional scrutiny. Set blocking rules for OWASP Top 10 2021 categories in AI-generated commits.
Train Reviewers on AI Code Patterns
Your code reviewers need to recognize how AI generators implement framework features. Schedule monthly sessions reviewing real examples of AI-generated vulnerabilities in your frameworks. Focus on the gap between "code that works" and "code that's secure."
Create Framework-Specific Prompts
When your developers use AI coding assistants, provide them with security-focused prompt templates: "Generate authentication middleware using [Framework]'s secure session management, implementing OWASP ASVS v4.0.3 requirement 3.2.1." This constrains the AI toward secure patterns.
Establish AI Code Acceptance Criteria
Define what makes AI-generated code acceptable for your codebase. Include: passes SAST scan, uses framework security features correctly, includes input validation, implements proper error handling. Reject code that doesn't meet all criteria, regardless of functionality.
Test Framework Defaults
Don't assume your framework's defaults are secure. AI generators will use defaults unless explicitly directed otherwise. Audit your framework configurations against NIST 800-53 Rev 5 Control CM-6 (Configuration Settings). Document required deviations from defaults.
Version Control Your AI Prompts
Store your security-focused AI prompts in your repository. Treat them like infrastructure as code. Review and update them when you discover new vulnerability patterns. This creates institutional knowledge about safe AI code generation.
The 15-vulnerability baseline isn't inevitable. It's the cost of treating AI generators as neutral tools rather than framework-dependent risk sources. Your framework choices create the vulnerability surface. Your security controls determine whether those vulnerabilities reach production.
