{"id":26939,"date":"2026-09-27T19:57:23","date_gmt":"2026-09-27T17:57:23","guid":{"rendered":"https:\/\/www.itspy.cz\/thesis\/nova-obranna-metoda-proti-jailbreak-utokum-na-velkejazykove-modely\/"},"modified":"2026-09-27T20:08:34","modified_gmt":"2026-09-27T18:08:34","slug":"nova-obranna-metoda-proti-jailbreak-utokum-na-velkejazykove-modely","status":"publish","type":"thesis","link":"https:\/\/www.itspy.cz\/sk\/thesis\/nova-obranna-metoda-proti-jailbreak-utokum-na-velkejazykove-modely\/","title":{"rendered":"Nov\u00e1 obrann\u00e1 metoda proti jailbreak \u00fatok\u016fm na velk\u00e9\njazykov\u00e9 modely"},"content":{"rendered":"<p>I bezpe\u010dnostn\u00ed pravidla modern\u00edch AI asistent\u016f lze obej\u00edt.<br \/>\nPomoc\u00ed tzv. jailbreak \u00fatok\u016f m\u016f\u017ee \u00fato\u010dn\u00edk p\u0159im\u011bt jazykov\u00fd<br \/>\nmodel, aby odpov\u011bd\u011bl na po\u017eadavky, kter\u00e9 by za b\u011b\u017en\u00fdch<br \/>\nokolnost\u00ed odm\u00edtl. Navr\u017een\u00e1 metoda AlcaTRAz proto funguje<br \/>\njako ochrann\u00e1 vrstva p\u0159ed samotn\u00fdm modelem. Z p\u0159\u00edklad\u016f<br \/>\n\u0161kodliv\u00fdch \u00fatok\u016f se automaticky nau\u010d\u00ed pravidlo, podle<br \/>\nkter\u00e9ho n\u00e1sledn\u011b c\u00edlen\u011b upravuje vybran\u00e1 m\u00edsta v textu. T\u00edm<br \/>\nnaru\u0161uje strukturu jailbreak \u00fatok\u016f, zat\u00edmco v\u00fdznam b\u011b\u017en\u00fdch<br \/>\ndotaz\u016f z\u016fst\u00e1v\u00e1 zachov\u00e1n.<\/p>\n<p>Metoda byla otestov\u00e1na na 22 typech jailbreak \u00fatok\u016f a 33<br \/>\njazykov\u00fdch modelech. V\u00fdsledky uk\u00e1zaly v\u00fdrazn\u00e9 omezen\u00ed<br \/>\nschopnosti model\u016f plnit \u0161kodliv\u00e9 po\u017eadavky, zat\u00edmco jejich<br \/>\nschopnost odpov\u00eddat na b\u011b\u017en\u00e9 dotazy z\u016fstala t\u00e9m\u011b\u0159 zachov\u00e1na.<\/p>\n<p>AlcaTRAz p\u0159itom nevy\u017eaduje p\u0159\u00edstup k vnit\u0159n\u00edm \u010d\u00e1stem modelu<br \/>\nani jeho p\u0159etr\u00e9nov\u00e1n\u00ed. V budoucnu by tak mohl slou\u017eit jako<br \/>\nlehk\u00e1 bezpe\u010dnostn\u00ed vrstva pro AI asistenty, firemn\u00ed<br \/>\nchatboty nebo dal\u0161\u00ed slu\u017eby zalo\u017een\u00e9 na jazykov\u00fdch modelech.<\/p>\n","protected":false},"featured_media":26751,"template":"","meta":{"_acf_changed":false,"_links_to":"","_links_to_target":""},"university":[190,188],"thesis-year":[468],"class_list":["post-26939","thesis","type-thesis","status-publish","has-post-thumbnail","hentry","thesis-year-468"],"acf":{"autor":"Petr Ka\u0161ka","portret":"","vedouci":"Ing. Jakub Re\u0161"},"_links":{"self":[{"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/thesis\/26939","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/thesis"}],"about":[{"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/types\/thesis"}],"version-history":[{"count":1,"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/thesis\/26939\/revisions"}],"predecessor-version":[{"id":27104,"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/thesis\/26939\/revisions\/27104"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/media\/26751"}],"wp:attachment":[{"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/media?parent=26939"}],"wp:term":[{"taxonomy":"university","embeddable":true,"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/university?post=26939"},{"taxonomy":"thesis-year","embeddable":true,"href":"https:\/\/www.itspy.cz\/sk\/wp-json\/wp\/v2\/thesis-year?post=26939"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}