{"id":"W4411112895","doi":"10.18653/v1/2025.findings-naacl.104","title":"Can’t Hide Behind the API: Stealing Black-Box Commercial Embedding Models","year":2025,"lang":"en","type":"article","venue":"","topic":"Data Quality and Management","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Waterloo","funders":"Natural Sciences and Engineering Research Council of Canada; Compute Canada","keywords":"Computer science; Embedding; Black box; Computer security; Artificial intelligence","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007435059,0.001526666,0.0009939423,0.001174979,0.0004999311,0.002418421,0.002406377,0.001753358,0.004476562],"category_scores_gemma":[0.03773772,0.0006665724,0.001287539,0.001118243,0.001198298,0.007972261,0.003317338,0.003391736,0.003811066],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0008927921,"about_ca_system_score_gemma":0.0009910072,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004379815,"about_ca_topic_score_gemma":0.006486825,"domain_scores_codex":[0.9968249,0.001747908,0.0002191669,0.0005064853,0.0005023751,0.000199278],"domain_scores_gemma":[0.9792064,0.01233824,0.0007260224,0.006263448,0.001147785,0.0003180111],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.001616884,0.0009159597,0.01213993,0.0007142089,0.0004303646,0.0004323216,0.001002287,0.2906479,0.008822011,0.02430791,0.03729414,0.6216761],"study_design_scores_gemma":[0.0000621263,0.0001993955,0.0004060866,0.00004602745,0.00003160411,0.0001002272,0.0001245603,0.9749219,0.004393795,0.01561271,0.004070146,0.00003147726],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2173888,0.00166199,0.7537571,0.003250301,0.0004087362,0.0003240253,0.00168253,0.01537459,0.006152025],"genre_scores_gemma":[0.7017302,0.0005880794,0.2832041,0.001197604,0.0001800312,0.0002925577,0.004895802,0.001367026,0.006544766],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.007435059,"threshold_uncertainty_score":0.03932077,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.2138067145493294,"score_gpt":0.4509011484647079,"score_spread":0.2370944339153785,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}