AI Model Analyzes and Patches Raw Executables Without Disassembler or Source Code
Researchers George Xenos and Sotiris Giannopoulos gave the GPT-6.1 Sol model only a hex dump of compiled programs, with tools removed, and it correctly explained all four programs' behavior. It decoded the stripped FreeBSD ministat utility and found a statistical bug where the 99.5% confidence option used critical values for about 99.8%, a flaw later confirmed in the source code. In further tests, the model directly edited binaries to add features to three programs and wrote three working executables from specifications. It also identified the behavior of an unknown malware sample and extracted correct indicators of compromise. The findings suggest possible new products for US startups in reverse engineering and malware analysis, though reliability, customer trust and liability for flawed AI-generated binaries remain open questions.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in