{"id":113971,"date":"2024-02-29T18:25:41","date_gmt":"2024-02-29T16:25:42","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/v-repozitorii-hugging-face-vyyavleny-vredonosnye-ai-modeli-vypolnyayushhie-kod"},"modified":"2024-02-29T18:25:41","modified_gmt":"2024-02-29T16:25:42","slug":"v-repozitorii-hugging-face-vyyavleny-vredonosnye-ai-modeli-vypolnyayushhie-kod","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/v-repozitorii-hugging-face-vyyavleny-vredonosnye-ai-modeli-vypolnyayushhie-kod","title":{"rendered":"N\u00eb repozitorin Hugging Face jan\u00eb zbuluar modele t\u00eb d\u00ebmshme AI q\u00eb kryejn\u00eb kod","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Hulumtuesit nga kompania JFrog kan\u00eb zbuluar n\u00eb depozitat e Hugging Face modele t\u00eb d\u00ebmshme t\u00eb m\u00ebsimit makinerik, instaluar e cila mund t\u00eb \u00e7oj\u00eb n\u00eb ekzekutimin e kodit nga sulmuesi p\u00ebr t\u00eb fituar kontroll mbi sistemin e p\u00ebrdoruesit. Problemi buron nga fakti se disa formate shp\u00ebrndarjeje t\u00eb modeleve lejojn\u00eb p\u00ebrfshirjen e kodit ekzekutues, p\u00ebr shembull, modelet q\u00eb p\u00ebrdorin formatin \"pickle\" mund t\u00eb p\u00ebrmbajn\u00eb objekte t\u00eb serializuara n\u00eb gjuh\u00ebn Python, si dhe kod q\u00eb ekzekutohet gjat\u00eb ngarkimit t\u00eb skedarit, nd\u00ebrsa modelet Tensorflow Keras mund t\u00eb ekzekutojn\u00eb kod p\u00ebrmes Lambda Layer. <\/p>\n
P\u00ebr t\u00eb parandaluar shp\u00ebrndarjen e modeleve t\u00eb tilla t\u00eb d\u00ebmshme, Hugging Face aplikon skanimin p\u00ebr z\u00ebvend\u00ebsimin e kodit t\u00eb serializuar, por modelet e d\u00ebmshme t\u00eb identifikuara tregojn\u00eb se verifikimet e ekzistuese mund t\u00eb anashkalohen. P\u00ebr m\u00eb tep\u00ebr, Hugging Face n\u00eb shumic\u00ebn e rasteve thjesht i sh\u00ebnon modelet si t\u00eb rrezikshme, pa bllokuar qasjen n\u00eb to. N\u00eb total jan\u00eb identifikuar rreth 100 modele potencialisht t\u00eb d\u00ebmshme, 95% e t\u00eb cilave jan\u00eb destinuar p\u00ebr p\u00ebrdorim me korniz\u00ebn PyTorch, dhe 5% me Tensorflow. Ndryshimet e d\u00ebmshme m\u00eb t\u00eb zakonshme p\u00ebrfshijn\u00eb kapjen e objekteve, organizimin e hyrjes s\u00eb jashtme n\u00eb sistem (reverse shell), nisjen e aplikacioneve dhe shkrimin n\u00eb skedar.
<\/center> <\/p>\n
V\u00ebren se, nga veprimet e kryera, shumica e modeleve t\u00eb d\u00ebmshme t\u00eb identifikuara jan\u00eb krijuar nga hulumtues t\u00eb siguris\u00eb q\u00eb p\u00ebrpiqen t\u00eb fitojn\u00eb shp\u00ebrblime p\u00ebr zbulesat e vulnerabiliteteve dhe metodeve p\u00ebr t\u00eb anashkaluar mbrojtjen e Hugging Face (p\u00ebr shembull, n\u00eb vend t\u00eb nj\u00eb sulmi real, k\u00ebto modele p\u00ebrpiqen t\u00eb nisin nj\u00eb kalkulator ose t\u00eb d\u00ebrgojn\u00eb nj\u00eb k\u00ebrkes\u00eb rrjeti me informacion p\u00ebr suksesin e sulmit). Nd\u00ebrkoh\u00eb, ndodhin edhe raste q\u00eb nisin nj\u00eb shell t\u00eb kthyer p\u00ebr t\u00eb lidhur sulmuesin me sistemin. <\/p>\n
P\u00ebr shembull, modelet \"baller423\/goober2\" dhe \"star23\/baller13\" jan\u00eb t\u00eb fokusuar n\u00eb kryerjen e sulmit ndaj sistemeve q\u00eb ngarkojn\u00eb skedarin e modelit n\u00eb PyTorch p\u00ebrmes funksionit torch.load(). P\u00ebr t\u00eb organizuar ekzekutimin e kodit, p\u00ebrdoret metoda \"__reduce__\" nga moduli pickle, e cila lejon futjen e kodit t\u00eb rast\u00ebsish\u00ebm n\u00eb procesin e deserializimit q\u00eb ekzekutohet gjat\u00eb ngarkimit t\u00eb modelit. \n Burimi: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"
![\"N\u00eb](\"\/wp-content\/uploads\/2024\/02\/41044e0439dc5ab9dafa74966309f7e7.png\")
<\/center> <\/p>\n