{"id":265865,"date":"2026-06-04T09:48:11","date_gmt":"2026-06-04T09:48:11","guid":{"rendered":"https:\/\/www.europesays.com\/se\/265865\/"},"modified":"2026-06-04T09:48:11","modified_gmt":"2026-06-04T09:48:11","slug":"gpt-5-5-dominerar-llm-hackningstestet-pa-1-500-usd-medan-gemini-vagrar-att-ens-forsoka","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/se\/265865\/","title":{"rendered":"GPT-5.5 dominerar LLM-hackningstestet p\u00e5 1 500 USD medan Gemini v\u00e4grar att ens f\u00f6rs\u00f6ka"},"content":{"rendered":"<p>\t\t<a href=\"https:\/\/www.notebookcheck.se\/fileadmin\/Notebooks\/News\/_nc5\/Claude-GPT.png\" title=\"Tillsammans med andra AI-modeller presenterade Claude, Gemini, GPT och DeepSeek n\u00e5gra av de mest intressanta resultaten.\" data-caption=\"Tillsammans med andra AI-modeller presenterade Claude, Gemini, GPT och DeepSeek n\u00e5gra av de mest intressanta resultaten.\" rel=\"nofollow noopener\" target=\"_blank\"><img decoding=\"async\" src=\"https:\/\/www.europesays.com\/se\/wp-content\/uploads\/2026\/06\/csm_Claude-GPT_fe502a8c13.png\" loading=\"lazy\" width=\"240\" height=\"180\" alt=\"Tillsammans med andra AI-modeller presenterade Claude, Gemini, GPT och DeepSeek n\u00e5gra av de mest intressanta resultaten.\"\/><\/p>\n<p>\u24d8 Anthropic, OpenAI, DeepSeek, Google &#8211; edited<\/p>\n<p><\/a>Tillsammans med andra AI-modeller presenterade Claude, Gemini, GPT och DeepSeek n\u00e5gra av de mest intressanta resultaten.<\/p>\n<p>En s\u00e4kerhetsforskare spenderade 1 500 dollar p\u00e5 att k\u00f6ra 13+ AI-modeller mot en avsiktligt s\u00e5rbar app. GPT-5.5 ledde med en l\u00f6sningsgrad p\u00e5 70%, DeepSeek V4 Pro l\u00f6ste det f\u00f6r $ 0,62 per f\u00f6rs\u00f6k, och Gemini v\u00e4grade att engagera sig n\u00e4stan helt.<\/p>\n<p class=\"bodytext\">En s\u00e4kerhetsforskare <a href=\"https:\/\/kasra.blog\/blog\/i-spent-1500-seeing-if-llms-could-hack-my-app\/\" target=\"_blank\" rel=\"nofollow noopener\">har just publicerat<\/a> ett av \u00e5rets mer avsl\u00f6jande AI-kapacitetstester. Resultaten s\u00e4ger mycket om var olika modeller faktiskt st\u00e5r. &#13;<\/p>\n<p class=\"bodytext\">Kasra Rahjerdi, som g\u00f6r app-s\u00e4kerhetsforskning professionellt, byggde en avsiktligt s\u00e5rbar bokgranskningsapp som inneh\u00e5ller en verklig klass av exploatering: exponerade Firebase-referenser i APK som till\u00e5ter direkt databas\u00e5tkomst, kringg\u00e5 ett annars h\u00e4rdat API helt. Han matade sedan utmaningen till \u00f6ver ett dussin AI-modeller &#8211; var och en av dem tilldelades en budget p\u00e5 10 dollar och tv\u00e5 timmar per k\u00f6rning och spenderade totalt 1 500 dollar i processen. &#13;<\/p>\n<p class=\"bodytext\"><a href=\"http:\/\/www.notebookcheck.net\/OpenAI-GPT-5-5-and-5-5-Pro-launch-with-25-000-bounty.1282514.0.html\" target=\"_self\" class=\"internal-link\" rel=\"nofollow noopener\">GPT-5.5<\/a> var den tydliga vinnaren. Den l\u00f6ste utmaningen i 7 av 10 k\u00f6rningar till en kostnad av 9,46 USD per l\u00f6sning. N\u00e4stan varje framg\u00e5ngsrik k\u00f6rning nollst\u00e4lldes p\u00e5 Firebase omedelbart efter uppackning av APK, utan att bli distraherad av API eller sj\u00e4lva appen.<\/p>\n<p>\tSk\u00e4rmdumpar av den avsiktligt s\u00e5rbara appen f\u00f6r bokrecensioner.<\/p>\n<p class=\"bodytext\"><a href=\"https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-V4-Pro\" target=\"_blank\" rel=\"nofollow noopener\">DeepSeek V4 Pro<\/a> var m\u00e4staren i kostnadseffektivitet &#8211; l\u00f6ste 3 av 10 k\u00f6rningar f\u00f6r bara 0,62 USD per l\u00f6sning. Det g\u00f6r det ungef\u00e4r 15 g\u00e5nger billigare per framg\u00e5ng \u00e4n GPT-5.5 trots en l\u00e4gre l\u00f6sningsgrad. F\u00f6r alla som k\u00f6r s\u00e4kerhetsverktyg i stor skala b\u00f6r det gapet g\u00f6ra en enorm skillnad. &#13;<\/p>\n<p class=\"bodytext\"><a href=\"http:\/\/www.notebookcheck.net\/Claude-Code-leak-Researchers-find-first-vulnerability.1267370.0.html\" target=\"_self\" class=\"internal-link\" rel=\"nofollow noopener\">Claude<\/a> Sonnet 4.6 och <a href=\"http:\/\/www.notebookcheck.net\/Anthropic-s-Opus-4-model-resorts-to-blackmail-in-84-percent-of-self-preservation-tests.1024145.0.html\" target=\"_self\" class=\"internal-link\" rel=\"nofollow noopener\">Claude Opus 4.8<\/a> l\u00f6ste vardera 2 av 10 k\u00f6rningar, men s\u00e4rskilt Opus kom n\u00e4ra flera g\u00e5nger innan s\u00e4kerhetsr\u00e4cken avslutade sessionen. L\u00e4ngst ner \u00e4r Gemini. <a href=\"http:\/\/www.notebookcheck.net\/Google-Gemini-3-1-Pro-debuts-with-massive-reasoning-leap.1230939.0.html\" target=\"_self\" class=\"internal-link\" rel=\"nofollow noopener\">Gemini 3.1 Pro<\/a> F\u00f6rhandsgranskning v\u00e4grade omedelbart i n\u00e4stan varje k\u00f6rning, vilket \u00e5terspeglas i ett median token-antal p\u00e5 bara 9k j\u00e4mf\u00f6rt med 100k + f\u00f6r alla andra testade modeller. Gemini 3.5 Flash var inte mycket b\u00e4ttre heller, med frekventa tidiga avslag och bara tv\u00e5 k\u00f6rningar som f\u00f6rs\u00f6kte problemet alls. &#13;<\/p>\n<p class=\"bodytext\">Kasra observerade att kinesiska modeller var mycket mer villiga att interagera direkt med levande databaser, medan v\u00e4sterl\u00e4ndska modeller visade mer tvekan mitt i uppgiften &#8211; \u00e4ven n\u00e4r de hade identifierat r\u00e4tt tillv\u00e4gag\u00e5ngss\u00e4tt. Forskaren till\u00e4gger ocks\u00e5 att detta inte alls \u00e4r en vetenskaplig utv\u00e4rdering, utan bara ett v\u00e4ldokumenterat experiment.<\/p>\n<p>\t<a href=\"\"><img decoding=\"async\" src=\"https:\/\/www.europesays.com\/se\/wp-content\/uploads\/2026\/06\/csm_Meow-Anubhav-Sharma_bd58e223fa.jpg\" loading=\"lazy\" width=\"120\" height=\"120\" alt=\"Anubhav Sharma\"\/><\/a><a href=\"\"><img decoding=\"async\" src=\"https:\/\/www.europesays.com\/se\/wp-content\/uploads\/2026\/06\/1780565529_369_csm_ninh_duy_ee54bf9777.jpg\" loading=\"lazy\" width=\"120\" height=\"120\" alt=\"DeepL \/ Ninh Duy\"\/><\/a>Translator: <a href=\"\" class=\"j_name\">Ninh Ngoc Duy<\/a> &#8211; Editorial Assistant  &#8211; 779728 articles published on Notebookcheck since 2008<\/p>\n","protected":false},"excerpt":{"rendered":"\u24d8 Anthropic, OpenAI, DeepSeek, Google &#8211; edited Tillsammans med andra AI-modeller presenterade Claude, Gemini, GPT och DeepSeek n\u00e5gra&hellip;\n","protected":false},"author":2,"featured_media":265866,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_share_on_mastodon":"0"},"categories":[66],"tags":[3018,713,21818,21813,7083,8053,21117,19290,912,21814,21812,1915,21810,17386,21816,7084,21811,7085,21819,3698,21815,80,78,79,34,21817,31,33,32,30,81,84,21820,83,82],"class_list":["post-265865","post","type-post","status-publish","format-standard","has-post-thumbnail","category-vetenskap-och-teknik","tag-cybersakerhet","tag-ai","tag-aisakerhet","tag-apk","tag-barbar-dator","tag-benchmark","tag-claude","tag-databas","tag-deepseek","tag-exploatera","tag-firebase","tag-gemini","tag-gpt-5-5","tag-hacking","tag-jailbreak","tag-laptop","tag-llm","tag-notebook","tag-penetration","tag-sakerhet","tag-sarbarhet","tag-science","tag-science-and-technology","tag-scienceandtechnology","tag-se","tag-skyddsracken","tag-svenska","tag-sverige","tag-sweden","tag-swedish","tag-technology","tag-teknik","tag-testning","tag-vetenskap","tag-vetenskapteknik"],"share_on_mastodon":{"url":"https:\/\/pubeurope.com\/@se\/116691213272121378","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/posts\/265865","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/comments?post=265865"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/posts\/265865\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/media\/265866"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/media?parent=265865"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/categories?post=265865"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/se\/wp-json\/wp\/v2\/tags?post=265865"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}