Citing issues with logic, correctness, and security, a new report recommends specific guardrails for AI-generated code.