Zinc-finger nucleases (ZFNs), transcription activator-like endonucleases (TALENs), and CRISPR-associated Cas9 endonucleases are three major generations of genome editing tools. However, no parallel comparison about the efficiencies and off-target activity of the three nucleases has been reported, which is critical for the final clinical decision. We for the first time developed the genome-wide unbiased identification of double-stranded breaks enabled by sequencing (GUIDE-seq) method in ZFNs and TALENs with novel bioinformatics algorithms to evaluate the off-targets. By targeting human papillomavirus 16 (HPV16), we compared the performance of ZFNs, TALENs, and SpCas9 in vivo. Our data showed that ZFNs with similar targets could generate distinct massive off-targets (287–1,856), and the specificity could be reversely correlated with the counts of middle “G” in zinc finger proteins (ZFPs). We also compared the TALENs with different N-terminal domains (wild-type [WT]/αN/βN) and G recognition modules (NN/NH) and found the design (αN or NN) to improve the efficiency of TALEN inevitably increased off-targets. Finally, our results showed that SpCas9 was more efficient and specific than ZFNs and TALENs. Specifically, SpCas9 had fewer off-target counts in URR (SpCas9, n = 0; TALEN, n = 1; ZFN, n = 287), E6 (SpCas9, n = 0; TALEN, n = 7), and E7 (SpCas9, n = 4; TALEN, n = 36). Taken together, we suggest that for HPV gene therapies, SpCas9 is a more efficient and safer genome editing tool. Our off-target data could be used to improve the design of ZFNs and TALENs, and the universal in vivo off-target detection pipeline for three generations of artificial nucleases provided useful tools for genome engineering-based gene therapy.