{"id":"W7106819272","doi":"10.48448/4nxg-ke53","title":"Jailbreak-Tuning: Models Efficiently Learn Jailbreak Susceptibility","year":2025,"lang":"","type":"other","venue":"Open MIND","topic":"","field":"","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"McGill University","funders":"","keywords":"USable; Component (thermodynamics); Work (physics); Frontier; Key (lock)","routes":{"ca_aff":true,"ca_fund":false,"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":[0.001900997,0.00163501,0.0008016216,0.0008926283,0.0005962544,0.00209446,0.002465906,0.00185348,0.007788212],"category_scores_gemma":[0.01565607,0.0008482044,0.001254605,0.0003739977,0.001042303,0.003316586,0.002353026,0.003338509,0.001911852],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001109747,"about_ca_system_score_gemma":0.001779148,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004722828,"about_ca_topic_score_gemma":0.008504123,"domain_scores_codex":[0.9992999,0.000240986,0.0000355903,0.0002043948,0.000118232,0.0001008989],"domain_scores_gemma":[0.9952989,0.002895012,0.0002759085,0.0009121984,0.0003617194,0.0002562706],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0003363343,0.0001996778,0.008213951,0.0002364828,0.000118605,0.0001851594,0.0003715043,0.8551359,0.00267588,0.02824108,0.01525128,0.08903418],"study_design_scores_gemma":[0.00002123451,0.00003450316,0.0001187765,0.00001993014,0.00001205744,0.00001756507,0.00002753559,0.975825,0.0006258499,0.02135093,0.001936469,0.00001016605],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.1147743,0.0007305679,0.8428191,0.002250577,0.0003319889,0.0002318158,0.001381473,0.02446162,0.01301851],"genre_scores_gemma":[0.7292962,0.0003405274,0.2557399,0.0009611201,0.0001244931,0.000403965,0.002742559,0.002878471,0.007512663],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.007788212,"threshold_uncertainty_score":0.02605414,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.06044998386102832,"score_gpt":0.3288474333551699,"score_spread":0.2683974494941416,"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."}}