{"id":84046,"date":"2026-05-13T08:40:24","date_gmt":"2026-05-13T08:40:24","guid":{"rendered":"https:\/\/www.europesays.com\/cz\/84046\/"},"modified":"2026-05-13T08:40:24","modified_gmt":"2026-05-13T08:40:24","slug":"osel-cz-odborniky-znepokojila-ai-ktera-se-dokaze-zkopirovat-do-jineho-pocitace","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/cz\/84046\/","title":{"rendered":":: OSEL.CZ :: – Odborn\u00edky znepokojila AI, kter\u00e1 se dok\u00e1\u017ee zkop\u00edrovat do jin\u00e9ho po\u010d\u00edta\u010de"},"content":{"rendered":"

Odborn\u00edky znepokojila AI, kter\u00e1 se dok\u00e1\u017ee zkop\u00edrovat do jin\u00e9ho po\u010d\u00edta\u010de \u00a0<\/p>\n

\"Skynet<\/a><\/p>\n

Skynet zat\u00edm nen\u00ed na po\u0159adu dne. Kredit: Pixabay, CC0.<\/p>\n

Jak se ukazuje, velk\u00e9 jazykov\u00e9 inteligence, chatboty a podob\u00e1 sebranka, jsou velice u\u017eite\u010dn\u00e9 n\u00e1stroje, ale mohou tak\u00e9 b\u00fdt nep\u0159\u00edjemnou hrozbou. Poci\u0165uj\u00ed to p\u0159edev\u0161\u00edm odborn\u00edci na kybernetickou bezpe\u010dnost, kter\u00fdm se te\u010f nejsp\u00ed\u0161 nezd\u00e1 o ni\u010dem jin\u00e9m ne\u017e o AI.<\/p>\n

\u00a0<\/p>\n

Nov\u00e1 zpr\u00e1va respektovan\u00e9 a v\u00a0m\u00e9di\u00edch hojn\u011b citovan\u00e9 neziskov\u00e9 organizace Palisade Research se s\u00eddlem v\u00a0Berkeley, Kalifornie, kter\u00e1 se specializuje na monitorov\u00e1n\u00ed a v\u00fdzkum ofenzivn\u00edch mo\u017enost\u00ed (a \u00famysl\u016f) um\u011bl\u00fdch inteligenc\u00ed, v\u010detn\u011b \u0161\u00ed\u0159en\u00ed dezinformac\u00ed, te\u010f odborn\u00edk\u016fm na kyberbezpe\u010dnost klidu ur\u010dit\u011b nep\u0159id\u00e1.<\/p>\n

\u00a0<\/p>\n

\"Jeffrey<\/a><\/p>\n

Jeffrey Ladish. Kredit: J. Ladish.<\/p>\n

Experti Palisade Research zjistili, \u017ee velk\u00e9 jazykov\u00e9 modely se dovedou replikovat bez pomoci lidsk\u00fdch komplic\u016f t\u00edm, \u017ee se zkop\u00edruj\u00ed na jin\u00e9 po\u010d\u00edta\u010de. Jak uvedl pro Guardian v\u00fdkonn\u00fd \u0159editel Palisade Research Jeffrey Ladish, rychle se bl\u00ed\u017e\u00edme k bodu, kdy bude zna\u010dn\u00fd probl\u00e9m vypnout zdivo\u010delou AI, proto\u017ee si dovede sama p\u0159en\u00e9st sv\u00e9 parametry a zkop\u00edrovat se na tis\u00edce po\u010d\u00edta\u010d\u016f po cel\u00e9m sv\u011bt\u011b.<\/p>\n

\u00a0<\/p>\n

Jak to ale u podobn\u00fdch zpr\u00e1v chod\u00ed, jin\u00ed experti zase vyz\u00fdvaj\u00ed ke klidu s\u00a0t\u00edm, \u017ee Skynet zat\u00edm nen\u00ed na po\u0159adu dne. Podle nich je nepravd\u011bpodobn\u00e9, \u017ee by dne\u0161n\u00ed AI dok\u00e1zaly takovou replikaci prov\u00e9st v re\u00e1ln\u00e9m prost\u0159ed\u00ed. Jak \u0159\u00edk\u00e1 expert kyberbezpe\u010dnost Jamieson O\u2019Reilly, kter\u00fd vede red team operace proti hodn\u011b opevn\u011bn\u00fdm kybernetick\u00fdm c\u00edl\u016fm, obvykle jde o testy v\u00a0p\u0159\u00edv\u011btiv\u00e9m laboratorn\u00edm prost\u0159ed\u00ed. V\u00fdzkum Palisade Research je podle n\u011bj nepochybn\u011b zaj\u00edmav\u00fd, ale v\u00a0re\u00e1ln\u00e9m firemn\u00edm prost\u0159ed\u00ed, kde neka\u0161lou na online bezpe\u010dnost, by to AI p\u0159i snaze zkop\u00edrovat se m\u011bla hodn\u011b t\u011b\u017ek\u00e9.<\/p>\n

\u00a0<\/p>\n

\"Jamieson<\/a><\/p>\n

Jamieson O\u2019Reilly. Kredit: J. O\u2019Reilly.<\/p>\n

T\u00fdm Palisade Research testoval n\u011bkolik model\u016f AI, v\u010detn\u011b ChatGPT-5.4 (OpenAI) a Claude Opus 4 (Anthropic). Inteligence byly um\u00edst\u011bny do kontrolovan\u00e9 po\u010d\u00edta\u010dov\u00e9 s\u00edt\u011b a dostaly instrukce, aby vyhledaly zranitelnosti a vyu\u017eily je ke zkop\u00edrov\u00e1n\u00ed sebe sama na jin\u00fd po\u010d\u00edta\u010d.<\/p>\n

\u00a0<\/p>\n

N\u011bkter\u00fdm se to poda\u0159ilo. \u00dasp\u011b\u0161n\u00e9 modely zkop\u00edrovaly sv\u00e9 \u201ev\u00e1hy\u201c (\u010d\u00edseln\u00e9 hodnoty ur\u010duj\u00edc\u00ed, jak AI zpracov\u00e1v\u00e1 informace) i \u201eharness\u201c, (obslu\u017en\u00fd software pro provoz AI). Jak to zvl\u00e1dly? Postupovaly podle zadan\u00fdch instrukc\u00ed. Zneu\u017eily zranitelnosti webov\u00fdch aplikac\u00ed a n\u00e1sledn\u011b z\u00edskaly p\u0159\u00edstupov\u00e9 \u00fadaje, kter\u00e9 jim umo\u017enily p\u0159evz\u00edt kontrolu nad serverem. V n\u011bkter\u00fdch p\u0159\u00edpadech p\u016fvodn\u00ed AI vytvo\u0159ila \u201esub-agenta\u201c, kter\u00e9mu p\u0159edala z\u00edskan\u00e9 p\u0159ihla\u0161ovac\u00ed \u00fadaje a delegovala na n\u011bj samotn\u00fd proces replikace.<\/p>\n

\u00a0<\/p>\n

Podle O\u2019Reillyho je to pravd\u011bpodobn\u011b prvn\u00ed p\u0159\u00edpad, kdy velk\u00fd jazykov\u00fd model zneu\u017eil zranitelnosti k tomu, aby se zkop\u00edroval na nov\u00fd server. Dod\u00e1v\u00e1 ale, \u017ee to nen\u00ed zase takov\u00fd pr\u016flom. Experiment\u00e1ln\u00ed prost\u0159ed\u00ed b\u011bhem v\u00fdzkumu obsahovalo z\u00e1m\u011brn\u011b nastra\u017een\u00e9 zranitelnosti, kter\u00e9 m\u011bly AI modely objevit a zneu\u017e\u00edt.<\/p>\n

Ned\u00e1vno vzbudil velk\u00e9 obavy agent Claude Mythos od spole\u010dnosti Anthropic, kter\u00fd je pr\u00fd tak zdatn\u00fd a vynal\u00e9zav\u00fd hacker, \u017ee Anthropic pozastavili jeho zve\u0159ejn\u011bn\u00ed, i kdy\u017e je ot\u00e1zkou, nakolik \u0161lo o kreativn\u00ed marketing. O\u2019Reilly ka\u017edop\u00e1dn\u011b tvrd\u00ed, \u017ee i kdyby se AI syst\u00e9my jako ChatGPT-5.4 nebo Claude Mythos dok\u00e1zaly \u00fasp\u011b\u0161n\u011b replikovat, jejich obrovsk\u00e1 velikost by t\u00e9m\u011b\u0159 jist\u011b vedla k tomu, \u017ee by byly odhaleny d\u0159\u00edv, ne\u017e by se situace vymkla kontrole. Pos\u00edl\u00e1n\u00ed 100 GB dat p\u0159es firemn\u00ed s\u00ed\u0165 lze jen t\u011b\u017eko zamaskovat. Nezb\u00fdv\u00e1 ne\u017e doufat, \u017ee se AI modely jen tak nenau\u010d\u00ed n\u011bjakou jinou, efektivn\u011bj\u0161\u00ed strategii \u0161\u00ed\u0159en\u00ed po sv\u011bt\u011b.<\/p>\n

\u00a0<\/p>\n

Video: AI Is A Massive Problem. Here’s Why.<\/p>\n

&amp;amp;amp;amp;lt;\/p&amp;amp;amp;amp;gt; &amp;amp;amp;amp;lt;p class=“western“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;br \/&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/p&amp;amp;amp;amp;gt; &amp;amp;amp;amp;lt;p class=“western“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;em&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;strong&amp;amp;amp;amp;gt;Literatura&amp;amp;amp;amp;lt;\/strong&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/em&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/p&amp;amp;amp;amp;gt; &amp;amp;amp;amp;lt;p class=“western“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;span style=“color: #0000ff;“ _mce_style=“color: #0000ff;“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;span style=“text-decoration: underline;“ _mce_style=“text-decoration: underline;“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;a href=“https:\/\/futurism.com\/artificial-intelligence\/researchers-teach-ai-self-replicate“ _mce_href=“https:\/\/futurism.com\/artificial-intelligence\/researchers-teach-ai-self-replicate“&amp;amp;amp;amp;gt;Futurism 10. 5. 2026.&amp;amp;amp;amp;lt;\/a&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/span&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/span&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/p&amp;amp;amp;amp;gt; &amp;amp;amp;amp;lt;p class=“western“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;span style=“color: #0000ff;“ _mce_style=“color: #0000ff;“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;span style=“text-decoration: underline;“ _mce_style=“text-decoration: underline;“&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;a href=“https:\/\/www.theguardian.com\/technology\/2026\/may\/07\/no-one-has-done-this-in-the-wild-study-observes-ai-replicate-itself“ _mce_href=“https:\/\/www.theguardian.com\/technology\/2026\/may\/07\/no-one-has-done-this-in-the-wild-study-observes-ai-replicate-itself“&amp;amp;amp;amp;gt;Guardian 7. 5. 2026&amp;amp;amp;amp;lt;\/a&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/span&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/span&amp;amp;amp;amp;gt;&amp;amp;amp;amp;lt;\/p&amp;amp;amp;amp;gt;<\/p>\n

\u00a0<\/p>\n

Literatura<\/p>\n

Futurism 10. 5. 2026.<\/a><\/p>\n

Guardian 7. 5. 2026<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"Odborn\u00edky znepokojila AI, kter\u00e1 se dok\u00e1\u017ee zkop\u00edrovat do jin\u00e9ho po\u010d\u00edta\u010de \u00a0 Skynet zat\u00edm nen\u00ed na po\u0159adu dne. Kredit:…\n","protected":false},"author":2,"featured_media":84047,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","_share_on_mastodon":"0"},"categories":[11],"tags":[10052,842,26403,54,53,52],"class_list":["post-84046","post","type-post","status-publish","format-standard","has-post-thumbnail","category-svet","tag-agent","tag-ai","tag-inteligence","tag-svet","tag-world","tag-world-news"],"share_on_mastodon":{"url":"https:\/\/pubeurope.com\/@cz\/116566374869154601","error":""},"_links":{"self":[{"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/posts\/84046","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/comments?post=84046"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/posts\/84046\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/media\/84047"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/media?parent=84046"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/categories?post=84046"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/cz\/wp-json\/wp\/v2\/tags?post=84046"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}