Jorge Gayoso

Static and LLM-Based Intra-Object Overflow Detection

Most buffer overflow detection tools (like AddressSanitizer) only check whether a write stays within an allocated object. This means they miss intra-object overflows — cases where a write stays within a struct but overflows past an individual field into neighboring fields. This thesis explores a combined approach: using CodeQL to statically identify suspicious overflow patterns in open-source C/C++ code, then using an LLM to generate inputs that dynamically confirm the bugs. The work also produces a curated dataset of real intra-object overflow instances, filling a gap in existing benchmarks.