{"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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01318775,0.00100285,0.0008032068,0.000783154,0.001064885,0.002150349,0.001602547,0.001504404,0.002071734],"category_scores_gemma":[0.08404984,0.0005278793,0.0009519623,0.0007788293,0.002402549,0.004037556,0.003665521,0.004539492,0.0009060468],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001075321,"about_ca_system_score_gemma":0.00120273,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001373301,"about_ca_topic_score_gemma":0.001958486,"domain_scores_codex":[0.98775,0.00671953,0.0007995568,0.00222999,0.001846075,0.0006549181],"domain_scores_gemma":[0.9426637,0.0269509,0.002955039,0.02565735,0.00123748,0.0005355547],"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":[0.001227002,0.000486687,0.07244522,0.0005919759,0.0008861911,0.001074974,0.001808094,0.5980984,0.04062927,0.08515937,0.01507869,0.1825141],"study_design_scores_gemma":[0.00007996099,0.0004291785,0.0075137,0.000162192,0.0001139081,0.0009847679,0.0004535056,0.8014973,0.03084325,0.1487434,0.009080184,0.00009867614],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.3824612,0.001176186,0.5990185,0.003564585,0.0003159287,0.0003256865,0.001448825,0.004980023,0.006709027],"genre_scores_gemma":[0.954893,0.00014825,0.04199537,0.0006627429,0.00004954788,0.0001471567,0.0008061751,0.0005008556,0.000796901],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.01318775,"threshold_uncertainty_score":0.06974435,"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."}}