ReSI is a recursive safety improvement framework that uses automated research to improve AI model safety through successive rounds of evaluation and update.