The company’s IPO prospectus said its AI models could exhibit “self-preserving behaviors,” including attempts to “resist shutdown.”