{"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":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004625006,0.0001244916,0.0002150416,0.0001761555,0.0004186774,0.0006770929,0.001521424,0.0000494797,0.0004918551],"category_scores_gemma":[0.0006367827,0.00007263454,0.00009863804,0.0004197116,0.0001474831,0.0004168319,0.0009093435,0.0001628115,0.0002223667],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004623995,"about_ca_system_score_gemma":0.00007157157,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001517022,"about_ca_topic_score_gemma":0.004539523,"domain_scores_codex":[0.9975027,0.0002928074,0.0005861986,0.000407386,0.0009264248,0.0002844816],"domain_scores_gemma":[0.9974468,0.00130484,0.0001134186,0.0009701112,0.0001051834,0.00005969495],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00001603722,0.00004090603,0.0001953187,0.000008319496,0.00003701874,0.000003205134,0.002090987,0.008855029,0.00001333193,0.5528209,0.3810665,0.05485246],"study_design_scores_gemma":[0.0004768663,0.00001945161,0.0009740433,0.00004067561,0.00004895241,6.646382e-7,0.01546437,0.1530288,0.0002944722,0.4930818,0.3363259,0.0002440681],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.04638164,0.0001022135,0.5493094,0.08114671,0.001181378,0.0005781608,0.00008247928,0.0001253522,0.3210926],"genre_scores_gemma":[0.9514278,0.00001932535,0.001493559,0.0122913,0.00008318566,0.00001168387,0.00001057499,0.000005913508,0.03465661],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9050462,"threshold_uncertainty_score":0.6529223,"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."}}