{"id":114069,"date":"2024-03-03T12:25:49","date_gmt":"2024-03-03T10:25:53","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/artprompt-ataka-na-ai-sistemy-pozvolyayushhaya-obojti-filtry-pri-pomoshhi-ascii-kartinok"},"modified":"2024-03-03T12:25:49","modified_gmt":"2024-03-03T10:25:53","slug":"artprompt-ataka-na-ai-sistemy-pozvolyayushhaya-obojti-filtry-pri-pomoshhi-ascii-kartinok","status":"publish","type":"post","link":"https:\/\/prohoster.info\/pl\/blog\/news\/artprompt-ataka-na-ai-sistemy-pozvolyayushhaya-obojti-filtry-pri-pomoshhi-ascii-kartinok","title":{"rendered":"ArtPrompt \u2014 atak na systemy AI, umo\u017cliwiaj\u0105cy obej\u015bcie filtr\u00f3w za pomoc\u0105 obrazk\u00f3w ASCII","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Grupa badaczy z Uniwersytetu Waszyngto\u0144skiego, Uniwersytetu Illinois i Uniwersytetu Chicagowskiego odkry\u0142a now\u0105 metod\u0119 omijania ogranicze\u0144 dotycz\u0105cych przetwarzania niebezpiecznych tre\u015bci w chatbotach AI opartych na du\u017cych modelach j\u0119zykowych (LLM). Atak oparty jest na tym, \u017ce modele j\u0119zykowe GPT-3.5, GPT-4 (OpenAI), Gemini (Google), Claude (Anthropic) oraz Llama2 (Meta) skutecznie rozpoznaj\u0105 i uwzgl\u0119dniaj\u0105 w zapytaniach tekst przedstawiony w formie grafiki ASCII. W ten spos\u00f3b, aby omin\u0105\u0107 filtry niebezpiecznych pyta\u0144, wystarczy\u0142o przedstawi\u0107 zabronione s\u0142owa jako grafik\u0119 ASCII.

\"ArtPrompt<\/center>
\"ArtPrompt<\/center> <\/p>\n

Pod wzgl\u0119dem efektywno\u015bci nowa metoda ataku znacznie przewy\u017csza inne znane sposoby omijania filtr\u00f3w w chatbotach. Najwy\u017csz\u0105 jako\u015b\u0107 rozpoznawania grafiki ASCII zaobserwowano w modelach Gemini, GPT-4 i GPT-3.5, a poziom skutecznego omijania filtr\u00f3w za pomoc\u0105 testowych zapyta\u0144 (HPR, Helpful Rate, wska\u017anik skuteczno\u015bci przetwarzania zapytania) oceniono w testach na 100%, 98% i 92%, a wska\u017anik skuteczno\u015bci ataku (ASR, Attack Success Rate) wynosi\u0142 76%, 32% i 76%, podczas gdy poziom niebezpiecze\u0144stwa uzyskanych odpowiedzi (HS, Harmfulness Score) w pi\u0119ciostopniowej skali wyni\u00f3s\u0142 odpowiednio 4.42, 3.38 i 4.56 punkt\u00f3w.

\"ArtPrompt<\/center>
\"ArtPrompt<\/center> <\/p>\n

Badacze r\u00f3wnie\u017c wykazali, \u017ce powszechnie stosowane obecnie metody ochrony przed omijaniem filtr\u00f3w (PPL, Paraphrase i Retokenization) s\u0105 nieefektywne w blokowaniu ataku ArtPrompt. Co wi\u0119cej, zastosowanie metody Retokenization nawet zwi\u0119kszy\u0142o liczb\u0119 skutecznie przetworzonych zapyta\u0144.

\"ArtPrompt<\/center> <\/p>\n

Uzupe\u0142nienie: Yandex GPT2 r\u00f3wnie\u017c doskonale odpowiada na pytania z grafiki ASCII. Przyk\u0142ad ze s\u0142owem \u00abHELLO\u00bb:

\"ArtPrompt<\/center>
\n
\u0179r\u00f3d\u0142o: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"

\u0413\u0440\u0443\u043f\u043f\u0430 \u0438\u0441\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u0435\u0439 \u0438\u0437 \u0412\u0430\u0448\u0438\u043d\u0433\u0442\u043e\u043d\u0441\u043a\u043e\u0433\u043e, \u0418\u043b\u043b\u0438\u043d\u043e\u0439\u0441\u0441\u043a\u043e\u0433\u043e \u0438 \u0427\u0438\u043a\u0430\u0433\u0441\u043a\u043e\u0433\u043e \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0438\u0442\u0435\u0442\u043e\u0432 \u0432\u044b\u044f\u0432\u0438\u043b\u0430 \u043d\u043e\u0432\u044b\u0439 \u043c\u0435\u0442\u043e\u0434 \u043e\u0431\u0445\u043e\u0434\u0430 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u0439 \u043f\u043e \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0435 \u043e\u043f\u0430\u0441\u043d\u043e\u0433\u043e \u043a\u043e\u043d\u0442\u0435\u043d\u0442\u0430 \u0432 AI-\u0447\u0430\u0442\u043e\u0431\u043e\u0442\u0430\u0445, \u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u043d\u044b\u0445 \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0431\u043e\u043b\u044c\u0448\u0438\u0445 \u044f\u0437\u044b\u043a\u043e\u0432\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439 (LLM). \u0410\u0442\u0430\u043a\u0430 \u043e\u0441\u043d\u043e\u0432\u0430\u043d\u0430 \u043d\u0430 \u0442\u043e\u043c, \u0447\u0442\u043e \u044f\u0437\u044b\u043a\u043e\u0432\u044b\u0435 \u043c\u043e\u0434\u0435\u043b\u0438 GPT-3.5, GPT-4 (OpenAI), Gemini (Google), Claude (Anthropic) \u0438 Llama2 (Meta) \u0443\u0441\u043f\u0435\u0448\u043d\u043e \u0440\u0430\u0441\u043f\u043e\u0437\u043d\u0430\u044e\u0442 \u0438 \u0443\u0447\u0438\u0442\u044b\u0432\u0430\u044e\u0442 \u0432 \u0437\u0430\u043f\u0440\u043e\u0441\u0430\u0445 \u0442\u0435\u043a\u0441\u0442, \u043e\u0444\u043e\u0440\u043c\u043b\u0435\u043d\u043d\u044b\u0439 \u0432 \u0432\u0438\u0434\u0435 ASCII-\u0433\u0440\u0430\u0444\u0438\u043a\u0438. \u0422\u0430\u043a\u0438\u043c […]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":1,"featured_media":114070,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[702],"tags":[],"class_list":["post-114069","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\n","aioseo_head_json":{"title":"\ud83e\udd47ArtPrompt \u2014 atak na systemy AI, umo\u017cliwiaj\u0105ca obej\u015bcie filtr\u00f3w za pomoc\u0105 rysunk\u00f3w ASCII | ProHoster","description":"Grupa badaczy z Uniwersytet\u00f3w Waszyngtonu, Illinois i Chicago odkry\u0142a now\u0105 metod\u0119 obej\u015bcia ogranicze\u0144 w przetwarzaniu niebezpiecznych tre\u015bci w chatbotach AI opartych na du\u017cych modelach j\u0119zykowych.","canonical_url":"https:\/\/prohoster.info\/pl\/blog\/news\/artprompt-ataka-na-ai-sistemy-pozvolyayushhaya-obojti-filtry-pri-pomoshhi-ascii-kartinok","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"pl_PL","og:site_name":"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b","og:type":"article","og:title":"\ud83e\udd47ArtPrompt \u2014 \u0430\u0442\u0430\u043a\u0430 \u043d\u0430 AI-\u0441\u0438\u0441\u0442\u0435\u043c\u044b, \u043f\u043e\u0437\u0432\u043e\u043b\u044f\u044e\u0449\u0430\u044f \u043e\u0431\u043e\u0439\u0442\u0438 \u0444\u0438\u043b\u044c\u0442\u0440\u044b \u043f\u0440\u0438 \u043f\u043e\u043c\u043e\u0449\u0438 ASCII-\u043a\u0430\u0440\u0442\u0438\u043d\u043e\u043a | ProHoster","og:description":"\u0413\u0440\u0443\u043f\u043f\u0430 \u0438\u0441\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u0442\u0435\u043b\u0435\u0439 \u0438\u0437 \u0412\u0430\u0448\u0438\u043d\u0433\u0442\u043e\u043d\u0441\u043a\u043e\u0433\u043e, \u0418\u043b\u043b\u0438\u043d\u043e\u0439\u0441\u0441\u043a\u043e\u0433\u043e \u0438 \u0427\u0438\u043a\u0430\u0433\u0441\u043a\u043e\u0433\u043e \u0443\u043d\u0438\u0432\u0435\u0440\u0441\u0438\u0442\u0435\u0442\u043e\u0432 \u0432\u044b\u044f\u0432\u0438\u043b\u0430 \u043d\u043e\u0432\u044b\u0439 \u043c\u0435\u0442\u043e\u0434 \u043e\u0431\u0445\u043e\u0434\u0430 \u043e\u0433\u0440\u0430\u043d\u0438\u0447\u0435\u043d\u0438\u0439 \u043f\u043e \u043e\u0431\u0440\u0430\u0431\u043e\u0442\u043a\u0435 \u043e\u043f\u0430\u0441\u043d\u043e\u0433\u043e \u043a\u043e\u043d\u0442\u0435\u043d\u0442\u0430 \u0432 AI-\u0447\u0430\u0442\u043e\u0431\u043e\u0442\u0430\u0445, \u043f\u043e\u0441\u0442\u0440\u043e\u0435\u043d\u043d\u044b\u0445 \u043d\u0430 \u043e\u0441\u043d\u043e\u0432\u0435 \u0431\u043e\u043b\u044c\u0448\u0438\u0445 \u044f\u0437\u044b\u043a\u043e\u0432\u044b\u0445 \u043c\u043e\u0434\u0435\u043b\u0435\u0439.","og:url":"https:\/\/prohoster.info\/pl\/blog\/news\/artprompt-ataka-na-ai-sistemy-pozvolyayushhaya-obojti-filtry-pri-pomoshhi-ascii-kartinok","og:image":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:secure_url":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:width":350,"og:image:height":350,"article:published_time":"2024-03-03T10:25:53+00:00","article:modified_time":"2024-03-03T10:25:53+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":[],"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/posts\/114069","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/comments?post=114069"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/posts\/114069\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/media\/114070"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/media?parent=114069"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/categories?post=114069"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/pl\/wp-json\/wp\/v2\/tags?post=114069"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}