Part III—Security Gauntlet

Reflecting on risks specific to AI-generated code, we define a mandatory multi-stage bulletproof vest.

Stage 1: Baseline (SAST & SCA)

• OWASP Top 10 pattern detection
• Dependency CVE scan
• Hardcoded secrets detection
• Input data validation and sanitization
• Application of strong encryption algorithms and key management
• Detection of dangerous functions/patterns (SQL string concatenation, eval, etc.)
• Proper authentication and session management
• Detection of missing authorization/access control
• Management of library vulnerabilities (Old versions or known vulns)

Stage 2: LLM-Specific Vulnerabilities

• Mimicry of Insecure Patterns
• False Confidence (Verification needed even with high confidence)
• Prompt Injection (Review prompt as first-class artifact)

Attack Vector	Mitigation (Example)
Instruction Override	Fix priority interpretation in system prompt
Roleplay / DAN	Prevent jailbreak via output validation
Obfuscation / Typo	Input normalization / Suspicious pattern detection
Template Destruction	Replace with harmless template prompt
Context Poisoning	Session validation before important items
Logic Manipulation	Make HITL human judgment mandatory

Stage 3: Supply Chain / Model Integrity

• Training data poisoning, algorithmic/ethical bias
• Model Due Diligence (Transparency reports / Security evaluation)