{"id":"W4417506422","doi":"10.1001/jamanetworkopen.2025.49963","title":"Vulnerability of Large Language Models to Prompt Injection When Providing Medical Advice","year":2025,"lang":"en","type":"article","venue":"JAMA Network Open","topic":"Adversarial Robustness in Machine Learning","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Vulnerability (computing); Adversarial system; Vulnerability assessment; Robustness (evolution); Quality (philosophy); Medical advice; Risk assessment; Language model","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":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004140352,0.0001503688,0.0003213098,0.00008913901,0.000234652,0.0002552193,0.002627051,0.0001573202,0.00007650864],"category_scores_gemma":[0.0007965117,0.0001404486,0.00005382208,0.0008940744,0.00002709664,0.001122254,0.003739112,0.0004974296,0.00001026633],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001038531,"about_ca_system_score_gemma":0.0003018819,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000890935,"about_ca_topic_score_gemma":0.0003111076,"domain_scores_codex":[0.9976117,0.0005512405,0.0004198008,0.0005510618,0.0004661373,0.0004000849],"domain_scores_gemma":[0.9984799,0.0003588736,0.0001527624,0.0007560238,0.0001242902,0.0001281296],"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":[0.0001842759,0.00019124,0.009325906,0.0001293672,0.00007823745,0.00002148667,0.008419857,0.6614897,0.00004130018,0.1914325,0.01904914,0.109637],"study_design_scores_gemma":[0.0006963161,0.00007152751,0.001857331,0.0003557995,0.000009820022,0.000003639254,0.0001830773,0.9779863,0.00008298396,0.0156408,0.002944981,0.0001674614],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.02465516,0.00007923308,0.9500214,0.005455284,0.0005773231,0.0008873167,0.000001634634,0.0001811069,0.01814156],"genre_scores_gemma":[0.8775595,0.000001759737,0.1191939,0.001820935,0.0002738827,0.00008322848,0.000005010712,0.00001180121,0.001050026],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8529043,"threshold_uncertainty_score":0.5727327,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.0176346959177542,"score_gpt":0.3256244953378211,"score_spread":0.3079897994200669,"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."}}