Updated
Updated · Forbes · Aug 13
Founders Get 5 Checks to Secure AI Apps as 45% of Model Tests Show Flaws
Updated
Updated · Forbes · Aug 13

Founders Get 5 Checks to Secure AI Apps as 45% of Model Tests Show Flaws

3 articles · Updated · Forbes · Aug 13

Summary

  • Five checks anchor the guidance for founders shipping AI-generated apps: review every line, run security scans, test access controls, remove embedded secrets, and require automated tests before release.
  • 45% of tests across code from more than 100 AI models produced risky security flaws, while Stack Overflow found 66% of developers called AI output “almost right, but not quite.”
  • Access and credential failures are highlighted as the most dangerous gaps: one startup exposed 1.5 million authentication tokens and 35,000 email addresses within three days, and another leak exposed a 512,000-line codebase.
  • The checklist comes as AI coding becomes mainstream, with an estimated 256 billion lines generated in 2024 and AI writing 41% to 46% of new production code.
  • Quality risks extend beyond security: Uplevel data cited in the report showed a 41% rise in bug rates after AI adoption, with no improvement in delivery time.

Insights

If AI writes half of all new code, who takes the blame when a hallucinated dependency triggers a massive data breach?
Are developers actually saving time with AI coding, or just trading writing hours for endless debugging of hidden security flaws?
Could the very tools designed to accelerate startup growth be silently planting the seeds of their inevitable collapse?