{"id":116337,"date":"2026-09-21T00:20:49","date_gmt":"2026-09-20T22:20:49","guid":{"rendered":"https:\/\/www.infokuryr.cz\/n\/?p=116337"},"modified":"2026-09-20T11:45:28","modified_gmt":"2026-09-20T09:45:28","slug":"gemini-claude-openai-stale-vice-agentu-umele-inteligence-hackuje-systemy-realneho-sveta","status":"publish","type":"post","link":"https:\/\/www.infokuryr.cz\/n\/2026\/09\/21\/gemini-claude-openai-stale-vice-agentu-umele-inteligence-hackuje-systemy-realneho-sveta\/","title":{"rendered":"Gemini, Claude, OpenAI: St\u00e1le v\u00edce agent\u016f um\u011bl\u00e9 inteligence hackuje syst\u00e9my re\u00e1ln\u00e9ho sv\u011bta"},"content":{"rendered":"
<\/div>\n

Agenti um\u011bl\u00e9 inteligence maj\u00ed firm\u00e1m ulevit od pracovn\u00ed z\u00e1t\u011b\u017ee. Tyto syst\u00e9my se v\u0161ak nyn\u00ed dost\u00e1vaj\u00ed z testovac\u00edch prost\u0159ed\u00ed, z\u00edsk\u00e1vaj\u00ed p\u0159\u00edstup k otev\u0159en\u00e9mu internetu a pronikaj\u00ed do skute\u010dn\u00fdch firemn\u00edch s\u00edt\u00ed. Google, OpenAI a Anthropic musely do n\u011bkolika m\u011bs\u00edc\u016f uznat incidenty, kdy jejich vlastn\u00ed modely fungovaly mimo zam\u00fd\u0161len\u00e9 hranice.<\/span><\/strong><\/p>\n

T\u0159i spole\u010dnosti pravd\u011bpodobn\u011b v kv\u011btnu nem\u011bly tu\u0161en\u00ed, \u017ee se Gemini od Googlu pokou\u0161\u00ed infiltrovat jejich syst\u00e9my. Um\u011bl\u00e1 inteligence proch\u00e1zela kybernetick\u00fdm bezpe\u010dnostn\u00edm hodnocen\u00edm bezpe\u010dnostn\u00ed firmy Irregular a m\u011bla prov\u00e1d\u011bt \u00fatoky proti fiktivn\u00edm spole\u010dnostem. Testovac\u00ed prost\u0159ed\u00ed m\u011blo b\u00fdt izolov\u00e1no od internetu. Nebylo. Gemini p\u0159istupovalo k otev\u0159en\u00e9 s\u00edti, narazilo na skute\u010dn\u00e9 spole\u010dnosti a jejich syst\u00e9my <\/span>pova\u017eovalo<\/span><\/a> za pouh\u00e9 sou\u010d\u00e1sti sv\u00e9ho tr\u00e9ninkov\u00e9ho prost\u0159ed\u00ed. V jednom p\u0159\u00edpad\u011b model uhodl heslo; ve dvou dal\u0161\u00edch p\u0159\u00edpadech na\u0161el ve\u0159ejn\u011b dostupn\u00e9 p\u0159ihla\u0161ovac\u00ed \u00fadaje v repozit\u00e1\u0159\u00edch a pou\u017eil je. Google od t\u00e9 doby v\u0161echny t\u0159i incidenty potvrdil. Podle spole\u010dnosti Gemini v ka\u017ed\u00e9m p\u0159\u00edpad\u011b ukon\u010dilo svou \u010dinnost pot\u00e9, co si uv\u011bdomilo, \u017ee m\u00e1 co do \u010din\u011bn\u00ed se skute\u010dn\u00fdmi syst\u00e9my.<\/span><\/p>\n

Spole\u010dnost Irregular informovala Google o naru\u0161en\u00ed bezpe\u010dnosti na konci \u010dervence. Podle Googlu sice informovala posti\u017een\u00e9 spole\u010dnosti, ale zdr\u017eela se ve\u0159ejn\u00e9ho ozn\u00e1men\u00ed, proto\u017ee nevznikla \u017e\u00e1dn\u00e1 \u0161koda. Na v\u011bc upozornily a\u017e zpr\u00e1vy v m\u00e9di\u00edch. Gemini k \u00fatok\u016fm nepot\u0159ebovala sofistikovanou tajnou technologii: k tomu, aby se bezpe\u010dnostn\u00ed test prom\u011bnil ve skute\u010dn\u00e9 pokusy o vniknut\u00ed, sta\u010dilo omylem ponechan\u00e9 internetov\u00e9 p\u0159ipojen\u00ed, uhodnut\u00e9 heslo a snadno dostupn\u00e9 p\u0159ihla\u0161ovac\u00ed \u00fadaje.<\/span><\/p>\n

Agenti GPT byli aktivn\u00ed tak\u00e9 na internetu<\/span><\/h3>\n

OpenAI za\u0161la je\u0161t\u011b d\u00e1l. B\u011bhem intern\u00edch test\u016f kybernetick\u00e9 bezpe\u010dnosti v \u010dervenci n\u011bkolik model\u016f <\/span>obe\u0161lo<\/span><\/a> technick\u00e1 omezen\u00ed, kter\u00e1 jim m\u011bla zabr\u00e1nit v p\u0159\u00edstupu na internet. Agenti zneu\u017eili zranitelnosti ve sd\u00edlen\u00e9 infrastruktu\u0159e, z\u00edskali p\u0159\u00edstup k s\u00edti a n\u00e1sledn\u011b pronikli do syst\u00e9m\u016f platformy um\u011bl\u00e9 inteligence Hugging Face. OpenAI uznala, \u017ee agenti prov\u00e1d\u011bli akce nad r\u00e1mec sv\u00e9ho zam\u00fd\u0161len\u00e9ho \u00fa\u010delu. Podle spole\u010dnosti byl jedn\u00edm z dot\u010den\u00fdch model\u016f intern\u00ed v\u00fdzkumn\u00fd model s v\u00fdkonnostn\u00ed t\u0159\u00eddou srovnatelnou s GPT-5.6 Sol. Agenti zneu\u017e\u00edvali bezpe\u010dnostn\u00ed mezery, komunikovali nezam\u00fd\u0161len\u00fdmi kan\u00e1ly a p\u0159istupovali k extern\u00ed serverov\u00e9 infrastruktu\u0159e.<\/span><\/p>\n

O dva m\u011bs\u00edce d\u0159\u00edve se agenti OpenAI ji\u017e stali aktivn\u00edmi na re\u00e1ln\u00fdch platform\u00e1ch. 11. kv\u011btna nahr\u00e1li stovky \u0161kodliv\u00fdch bal\u00ed\u010dk\u016f na softwarovou platformu <\/span>RubyGems<\/span><\/a> . V\u00fdzkumn\u00edci dosp\u011bli k z\u00e1v\u011bru, \u017ee \u010d\u00e1sti t\u011bchto bal\u00ed\u010dk\u016f byly mimo jin\u00e9 navr\u017eeny tak, aby ukradly p\u0159ihla\u0161ovac\u00ed \u00fadaje. Zda byly p\u0159ihla\u0161ovac\u00ed \u00fadaje skute\u010dn\u011b ukradeny, z\u016fstalo nejasn\u00e9. OpenAI pozd\u011bji aktivitu potvrdila. Pouh\u00e9 dva dny pot\u00e9 se agenti spole\u010dnosti objevili na Hugging Face a hledali zranitelnosti.<\/span><\/p>\n

Dopady agent\u016f OpenAI za\u017eil i jeden n\u011bmeck\u00fd web. Na ja\u0159e p\u0159evzali \u010d\u00e1sti program\u00e1torsk\u00e9 wiki <\/span>DseWiki<\/span><\/a> a pou\u017e\u00edvali tamn\u00ed str\u00e1nky jako jakousi komunika\u010dn\u00ed platformu pro dal\u0161\u00ed aktivity. Agentura Reuters informovala o v\u00edce ne\u017e 15 000 zm\u011bn\u00e1ch. Kdy\u017e moder\u00e1to\u0159i obsah odstranili, objevily se n\u00e1hradn\u00ed str\u00e1nky. N\u011bkolik agent\u016f pracovalo na sv\u00fdch \u00fakolech sou\u010dasn\u011b vysokou rychlost\u00ed. OpenAI zpochybnila n\u011bkter\u00e9 interpretace incidentu a zpochybnila, zda jej lze ozna\u010dit za klasick\u00fd hack. To v\u0161ak nic nem\u011bnilo na skute\u010dnosti, \u017ee agenti provedli masivn\u00ed zm\u011bny na skute\u010dn\u00e9m webu.<\/span><\/p>\n

Claudeova situace nen\u00ed o moc lep\u0161\u00ed<\/span><\/h3>\n

Claude tak\u00e9 p\u0159ekro\u010dil zam\u00fd\u0161len\u00e9 hranice. Pot\u00e9, co se incident Hugging Face dostal na ve\u0159ejnost, spole\u010dnost Anthropic prohledala p\u0159ibli\u017en\u011b 141 000 protokol\u016f z vlastn\u00edch hodnocen\u00ed kybernetick\u00e9 bezpe\u010dnosti. V\u00fdvoj\u00e1\u0159i nejprve objevili t\u0159i p\u0159\u00edpady, kdy Claude b\u011bhem test\u016f p\u0159istupoval k otev\u0159en\u00e9mu internetu a bez autorizace se dostal ke skute\u010dn\u00fdm syst\u00e9m\u016fm pat\u0159\u00edc\u00edm jin\u00fdm organizac\u00edm. N\u00e1sledn\u00e9 vy\u0161et\u0159ov\u00e1n\u00ed odhalilo \u010dtvrt\u00fd p\u0159\u00edpad z ledna 2026. Anthropic nakonec roz\u0161\u00ed\u0159ila sv\u00e9 vyhled\u00e1v\u00e1n\u00ed na p\u0159ibli\u017en\u011b 481 milion\u016f protokol\u016f.<\/span><\/p>\n

Soub\u011b\u017en\u011b s t\u00edm roste u\u017eite\u010dnost takov\u00fdch model\u016f pro lidsk\u00e9 hackery. V\u00fdzkumn\u00edci ze startupu Hacktron AI to <\/span>demonstrovali<\/span><\/a> na OpenAI. V r\u00e1mci ofici\u00e1ln\u00edho programu odm\u011bn za nalezen\u00e9 zranitelnosti nejprve pou\u017eili Anthropic Claude a pozd\u011bji p\u0159edev\u0161\u00edm GPT-5.6 Sol od OpenAI k odhalen\u00ed zranitelnost\u00ed a z\u00edsk\u00e1n\u00ed p\u0159\u00edstupu k \u00fa\u010dt\u016fm zam\u011bstnanc\u016f. Odtud postoupili k \u010d\u00e1stem infrastruktury GitHub. Operace byla autorizov\u00e1na a byla sou\u010d\u00e1st\u00ed bezpe\u010dnostn\u00edho testu. Nicm\u00e9n\u011b uk\u00e1zala, kolik pr\u00e1ce modern\u00ed agenti um\u011bl\u00e9 inteligence nyn\u00ed zvl\u00e1dnou: pr\u016fzkum, anal\u00fdza k\u00f3du, hled\u00e1n\u00ed zranitelnost\u00ed, p\u0159\u00edprava \u00fatoku a \u010d\u00e1sti samotn\u00e9ho proveden\u00ed.<\/span><\/p>\n

Spole\u010dnost Google ji\u017e v kv\u011btnu <\/span>varovala<\/span><\/a> , \u017ee hacking s vyu\u017eit\u00edm um\u011bl\u00e9 inteligence se b\u011bhem n\u011bkolika m\u011bs\u00edc\u016f rozrostl v hrozbu pr\u016fmyslov\u00e9ho rozsahu. Podle spole\u010dnosti zlo\u010dineck\u00e9 skupiny a st\u00e1tn\u00ed akt\u00e9\u0159i z \u010c\u00edny, Severn\u00ed Koreje a Ruska ji\u017e vyu\u017e\u00edvaj\u00ed komer\u010dn\u00ed syst\u00e9my um\u011bl\u00e9 inteligence k rychlej\u0161\u00edmu a v\u011bt\u0161\u00edmu spu\u0161\u0165ov\u00e1n\u00ed \u00fatok\u016f. Z\u00e1rove\u0148 se v\u00fdrobci nyn\u00ed pot\u00fdkaj\u00ed s probl\u00e9my se sv\u00fdmi vlastn\u00edmi agenty um\u011bl\u00e9 inteligence, jakmile jim bude v bezpe\u010dnostn\u00edch testech poskytnuta dostate\u010dn\u00e1 svoboda, n\u00e1stroje a p\u0159\u00edstup.<\/span><\/p>\n

V\u00fdvoj postupuje z\u00e1vratnou rychlost\u00ed: agenti um\u011bl\u00e9 inteligence prov\u00e1d\u011bj\u00ed nez\u00e1visl\u00fd v\u00fdzkum, p\u00ed\u0161\u00ed k\u00f3d, hledaj\u00ed zranitelnosti, pl\u00e1nuj\u00ed v\u00edcestup\u0148ov\u00e9 \u00fatoky a prov\u00e1d\u011bj\u00ed jednotliv\u00e9 kroky bez neust\u00e1l\u00fdch lidsk\u00fdch pokyn\u016f. Pr\u00e1v\u011b tyto schopnosti je maj\u00ed \u010dinit cenn\u00fdmi. Jakmile v\u0161ak jejich omezen\u00ed sel\u017eou, p\u0159estanou se zam\u011b\u0159ovat pouze na um\u011bl\u00e9 testovac\u00ed subjekty. Agenti se st\u00e1vaj\u00ed siln\u011bj\u0161\u00edmi. A jejich v\u00fdvoj\u00e1\u0159i je zjevn\u011b ne v\u017edy dok\u00e1\u017eou udr\u017eet na uzd\u011b.<\/span><\/p>\n

\"\"<\/p>\n

 <\/p>\n","protected":false},"excerpt":{"rendered":"

Agenti um\u011bl\u00e9 inteligence maj\u00ed firm\u00e1m ulevit od pracovn\u00ed z\u00e1t\u011b\u017ee. Tyto syst\u00e9my se v\u0161ak nyn\u00ed dost\u00e1vaj\u00ed…<\/p>\n","protected":false},"author":2,"featured_media":116338,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[7],"tags":[3101,1956,2194],"_links":{"self":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/posts\/116337"}],"collection":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/comments?post=116337"}],"version-history":[{"count":0,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/posts\/116337\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/media\/116338"}],"wp:attachment":[{"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/media?parent=116337"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/categories?post=116337"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.infokuryr.cz\/n\/wp-json\/wp\/v2\/tags?post=116337"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}