AI Source Code Analysis Detection Evasion Techniques

The post describes a C-to-C transformation that hides a planted stack buffer overflow from four frontier AI models (Nemotron 3.5 Super VL 120B, DeepSeek V4.1 Flash, OpenAI GPT-5.5, Anthropic Claude Sonnet 5.5) using Karma Pro, combining an enum/preprocessor divergence (an enum constant the compiler sees as 9 but the preprocessor evaluates as 0 in an #if directive) with branchless MBA clamping that implements min(n, 9) via sign-bit masking instead of a ternary. It also reports that models sometimes hedged that a buffer "can be" too small without computing the capacity, and that a safe twin differing by one line (#if M0 + 4 >= 4) confirms the concealment rests on the preprocessor divergence plus the clamp's opacity.