{"id":"W4417266690","doi":"10.48550/arxiv.2505.16789","title":"Accidental Vulnerability: Factors in Fine-Tuning that Shift Model Safeguards","year":2025,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Adversarial Robustness in Machine Learning","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Alliance de recherche numérique du Canada; University of Toronto; Government of Canada; Canadian Institute for Advanced Research","keywords":"Adversarial system; Interpretability; Robustness (evolution); Vulnerability (computing); Threat model; Key (lock); Accidental","routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":true},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006561935,0.0005937241,0.0006714531,0.0007917127,0.0002702538,0.0002131908,0.003746815,0.0005514473,0.00003742913],"category_scores_gemma":[0.0002011754,0.0007078781,0.0003412951,0.0009937818,0.0001326893,0.0009365084,0.007256817,0.002266058,0.00001486856],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0008666095,"about_ca_system_score_gemma":0.0006099538,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0009287252,"about_ca_topic_score_gemma":0.0006284748,"domain_scores_codex":[0.9964385,0.000432708,0.0003637619,0.001897204,0.0002310785,0.0006367671],"domain_scores_gemma":[0.9971869,0.0004092564,0.0003346784,0.001811768,0.00008146649,0.0001759315],"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":[0.00002123495,0.00005359546,0.1113045,0.00006610864,0.00003920549,0.0001394121,0.0008917103,0.8236918,0.000003816248,0.06344683,0.00002594387,0.0003158243],"study_design_scores_gemma":[0.0005306107,0.00001549824,0.01105066,0.0002002215,0.00004031088,9.571374e-7,0.0001503088,0.9372298,0.00005874737,0.05009447,0.00003176064,0.0005966833],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3666809,0.00002269503,0.6290584,0.0001226155,0.0006610112,0.0002548555,0.00001410666,0.000276038,0.002909377],"genre_scores_gemma":[0.9929567,0.00002455858,0.005432631,0.0000688651,0.00004760737,0.000001706461,0.00003105249,0.00002308968,0.001413784],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6262757,"threshold_uncertainty_score":0.9995372,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.07862829881679946,"score_gpt":0.2337805531565255,"score_spread":0.1551522543397261,"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."}}