SecTor 2026 | Using Claude or Codex? We Can Hack Your App?

Today, most code is actually vibe-coded by the same top AI code models, which often reuse the same implementation patterns, including vulnerability patterns. We found that an attacker can generate and inspect their own applications, discover vulnerabilities that developers may have missed, and use those findings to attack other applications generated by the same model, which have a high chance of containing the same vulnerabilities. In this Briefing, we evaluate how high these chances are among today’s top code models and present a new attack and threat model.

We introduce FSTab, a novel black-box attack that maps visible UI features such as login, upload, and password reset to recurring corresponding backend vulnerabilities. We evaluate how often these vulnerabilities transfer across applications, prompts, and domains using six widely used code LLMs. FSTab reaches up to 85.67% average attack success and 84.97% average vulnerability coverage, showing how model-level coding habits can expose vulnerabilities across thousands of AI-generated applications.

Amit LeVi | AI Researcher & Founder, ATLAS Lab
Raz Tel-Vered | AI and Security Research Team Lead, Zenity

https://blackhat.com/sector/briefings/schedule/#using-claude-or-codex-we-can-hack-your-app-extracting-recurring-vulnerabilities-from-black-box-llm-generated-software-55091Black HatRead More