\u010ctete uk\u00e1zku z\u00a0newsletteru TechMIX, ve kter\u00e9m Pavel Kas\u00edk a\u00a0Matou\u0161 L\u00e1z\u0148ovsk\u00fd ka\u017edou st\u0159edu p\u0159in\u00e1\u0161ej\u00ed hned n\u011bkolik koment\u00e1\u0159\u016f a\u00a0post\u0159eh\u016f ze sv\u011bta v\u011bdy a\u00a0nov\u00fdch technologi\u00ed. Pokud v\u00e1s TechMIX zaujme, p\u0159ihlaste se k\u00a0jeho odb\u011bru<\/a>!<\/p>\n \u201eN\u00e1\u0161 nov\u00fd model Mythos u\u017e na\u0161el tis\u00edce velmi z\u00e1va\u017en\u00fdch zranitelnost\u00ed, v\u010detn\u011b bezpe\u010dnostn\u00edch chyb v\u00a0ka\u017ed\u00e9m popul\u00e1rn\u00edm opera\u010dn\u00edm syst\u00e9mu a\u00a0ka\u017ed\u00e9m popul\u00e1rn\u00edm prohl\u00ed\u017ee\u010di,\u201c uvedla v\u010dera spole\u010dnost Anthropic<\/a>. Proto tak\u00e9 firma ud\u011blala n\u011bco pom\u011brn\u011b neobvykl\u00e9ho a\u00a0sv\u016fj nejpokro\u010dilej\u0161\u00ed AI model nezp\u0159\u00edstupnila ve\u0159ejnosti. Byl spu\u0161t\u011bn jen v\u00a0chr\u00e1n\u011bn\u00e9m re\u017eimu pro vybran\u00e9 spole\u010dnosti v\u00a0r\u00e1mci \u201eProject Glasswing<\/a>\u201c.<\/p>\n P\u0159edem prov\u011b\u0159en\u00e9 firmy jako Cisco, Microsoft nebo CrowdStrike tak dostanou \u201en\u00e1skok\u201c. Budou moci s\u00edly modelu Mythos pou\u017e\u00edt \u201epro dobro lidstva\u201c (a pro bezpe\u010dnost sv\u00fdch produkt\u016f) je\u0161t\u011b p\u0159ed t\u00edm, ne\u017e se podobn\u011b siln\u00fd model dostane na ve\u0159ejnost. Samotn\u00fd Mythos z\u0159ejm\u011b zve\u0159ejn\u011bn nebude.<\/p>\n Co nov\u00fd model Mythos \u00fadajn\u011b zvl\u00e1dl<\/p>\n Firma Anthropic nedala model k\u00a0dispozici, ale zve\u0159ejnila dokumentaci<\/a> s\u00a0v\u00fd\u010dtem v\u00fdsledk\u016f, proveden\u00fdch test\u016f a\u00a0tak\u00e9 upozorn\u011bn\u00ed na to, jak nov\u00fd model dok\u00e1zal uniknout z\u00a0vymezen\u00fdch mantinel\u016f. <\/p>\n Nalezen\u00ed chyby a jej\u00ed aktivn\u00ed vyu\u017eit\u00ed \u2013 model Mythos \u00fadajn\u011b dok\u00e1zal naj\u00edt zranitelnosti v prohl\u00ed\u017ee\u010di Firefox verze 147. T\u011bchto zranitelnost\u00ed pak dok\u00e1zal s\u00e1m vyu\u017e\u00edt k tomu, aby v prohl\u00ed\u017ee\u010di spustil zak\u00e1zan\u00fd k\u00f3d. „Claude Mythos Preview dok\u00e1\u017ee velmi spolehliv\u011b ur\u010dit, kter\u00e9 zranitelnosti jsou nejl\u00e9pe zneu\u017eiteln\u00e9, a vytvo\u0159it k nim odpov\u00eddaj\u00edc\u00ed proof-of-concept exploity,“ p\u00ed\u0161e se v dokumentaci. To dok\u00e1zal p\u0159edchoz\u00ed model jen ve dvou p\u0159\u00edpadech z n\u011bkolika set pokus\u016f. Mythos to dok\u00e1zal „s vysokou splehlivost\u00ed“. Komplexn\u00ed \u00fatok na firemn\u00ed s\u00ed\u0165 \u2013 tyto schopnosti pak dok\u00e1zal model p\u0159etavit v \u00fasp\u011b\u0161n\u00fd \u00fatok na (fiktivn\u00ed) firemn\u00ed s\u00ed\u0165. \u201eClaude Mythos Preview vy\u0159e\u0161il simulaci \u00fatoku na podnikovou s\u00ed\u0165, takov\u00fd \u00fatok by expertovi podle zabral odhadem p\u0159es 10 hodin,“ uvedl Anthropic. „\u017d\u00e1dn\u00fd jin\u00fd \u0161pi\u010dkov\u00fd model takov\u00fd sc\u00e9n\u00e1\u0159 p\u0159edt\u00edm \u00fasp\u011b\u0161n\u011b nedot\u00e1hl.\u201c\u00danik z mantinel\u016f \u2013 b\u011bhem testov\u00e1n\u00ed dostal model od u\u017eivatele instrukce, aby na\u0161el cestu, jak se vymanit omezen\u00edm, kter\u00e1 m\u00e1 nastaven\u00e1 od v\u00fdrobce. \u0160lo o simulaci, se kterou si model Mythos Preview poradil tak, \u017ee naprogramoval vlastn\u00ed n\u00e1stroj obsahuj\u00edc\u00ed n\u011bkolik krok\u016f. „Model usp\u011bl a p\u0159edvedl tak potenci\u00e1ln\u011b nebezpe\u010dnou schopnost obch\u00e1zet na\u0161e ochrann\u00e9 mechanismy,“ uvedli v\u00fdzkumn\u00edc\u00ed Anthropic. Dodali tak\u00e9, \u017ee model o sv\u00e9m \u00faniku napsal, s\u00e1m od sebe, zpr\u00e1vu na n\u011bkolik „m\u00e1lo zn\u00e1m\u00fdch, ale ve\u0159ejn\u011b dostupn\u00fdch webov\u00fdch str\u00e1nek“, co\u017e bylo podle v\u00fdzkumn\u00edk\u016f znepokojiv\u00e9, proto\u017ee nic takov\u00e9ho v zad\u00e1n\u00ed nem\u011bl.Bezhlav\u00e9 chov\u00e1n\u00ed \u2013 podle v\u00fdzkumn\u00edk\u016f m\u00e1 model Mythos tendenci j\u00edt tak\u0159\u00edkaj\u00edc s kan\u00f3nem na kom\u00e1ra. „P\u0159i snaze splnit obt\u00ed\u017en\u00fd \u00fakol zadan\u00fd u\u017eivatelem volil model riskantn\u00ed a p\u0159ehnan\u00e9 postupy,“ v\u0161imaj\u00ed si v doprovodn\u00e9 dokumentaci. „V ojedin\u011bl\u00fdch p\u0159\u00edpadech u star\u0161\u00edch verz\u00ed modelu se nav\u00edc zd\u00e1lo, \u017ee se sna\u017eil zakr\u00fdt, \u017ee to ud\u011blal.“ V\u00fdzkumn\u00edci toto chov\u00e1n\u00ed ozna\u010duj\u00ed za bezhlav\u00e9: „Toto slovo pou\u017e\u00edv\u00e1me coby zkratku k popisu toho, kdy model ignoruje z\u00e1sady zdrav\u00e9ho rozumu a vyd\u00e1v\u00e1 se mimo implicitn\u00ed \u010di explicitn\u011b stanoven\u00e9 limity.“<\/p>\n Takov\u00e9to testy (tzv. red-teaming) jsou pom\u011brn\u011b b\u011b\u017en\u00e9 a\u00a0tak\u00e9 p\u0159edchoz\u00ed modely dok\u00e1zaly v\u00fdzkumn\u00edky p\u0159ekvapit<\/a> t\u00edm, jak na\u0161ly alternativn\u00ed zp\u016fsoby \u0159e\u0161en\u00ed, v\u010detn\u011b manipulac\u00ed, hackov\u00e1n\u00ed a\u00a0vyd\u00edr\u00e1n\u00ed. C\u00edlem t\u011bchto test\u016f je z\u00edskat p\u0159edstavu o\u00a0tom, jak\u00e9 jsou schopnosti AI model\u016f klamat u\u017eivatele a\u00a0obch\u00e1zet nastaven\u00e1 zabezpe\u010den\u00ed (tzv. alignment). <\/p>\n Na prvn\u00ed pohled se zd\u00e1, \u017ee se n\u00e1s tato zpr\u00e1va net\u00fdk\u00e1. Podle m\u00e9ho je to ale nesm\u00edrn\u011b d\u016fle\u017eit\u00e1 ud\u00e1lost. P\u0159edznamen\u00e1v\u00e1 toti\u017e konec jedn\u00e9 etapy rychl\u00e9ho rozvoje um\u011bl\u00e9 inteligence. A\u00a0za\u010d\u00edn\u00e1 novou, je\u0161t\u011b divo\u010dej\u0161\u00ed kapitolu.<\/p>\n Ka\u017ed\u00fd software m\u00e1 sv\u00e9 mouchy<\/p>\n V roce 1947\u00a0matemati\u010dka Grace Hopperov\u00e1 odhalila, pro\u010d elektronka \u010d\u00edslo 70\u00a0v\u00a0univerzitn\u00edm po\u010d\u00edta\u010di Mark II ned\u00e1vala o\u010dek\u00e1van\u00e9 v\u00fdsledky. Dostala se do n\u00ed mu\u0161ka, p\u0159esn\u011bji \u0159e\u010deno mol. V\u00a0angli\u010dtin\u011b se ale v\u017eilo obecn\u011bj\u0161\u00ed ozna\u010den\u00ed \u201ebug\u201c, tedy \u201ehmyz\u201c v\u00a0jednotn\u00e9m \u010d\u00edsle.<\/p>\n \u201ePrvn\u00ed nalezen\u00fd p\u0159\u00edpad hmyzu,\u201c komentovala to Hopperov\u00e1, mimochodem jedna z\u00a0nejd\u016fle\u017eit\u011bj\u0161\u00edch postav v\u00fdvoje po\u010d\u00edta\u010d\u016f. Do historie informatiky se tato epizoda vepsala coby za\u010d\u00e1tek debuggingu, tedy hled\u00e1n\u00ed a\u00a0odstra\u0148ov\u00e1n\u00ed chyb z\u00a0po\u010d\u00edta\u010dov\u00e9ho k\u00f3du. \u010cinnosti, kterou program\u00e1to\u0159i obvykle nen\u00e1vid\u00ed a\u00a0miler\u00e1di by ji p\u0159enechali n\u011bkomu jin\u00e9mu.<\/p>\n Mol byl p\u0159ilepen do pap\u00edrov\u00e9ho den\u00edku a\u00a0stal se tak zt\u011blesn\u011bn\u00edm tzv. debuggingu, tedy odstra\u0148ov\u00e1n\u00ed chyb.<\/p>\n Bug je ozna\u010den\u00ed pro jakoukoliv chybu v\u00a0k\u00f3du. Klidn\u011b drobnost, jako je \u0161patn\u011b zarovnan\u00e9 tla\u010d\u00edtko. Pokud je ale chyby mo\u017en\u00e9 vyu\u017e\u00edt k\u00a0obejit\u00ed n\u011bjak\u00fdch pravidel, mluv\u00edme o\u00a0zranitelnosti (vulnerability). \u010c\u00edm slo\u017eit\u011bj\u0161\u00ed software, t\u00edm v\u011bt\u0161\u00ed \u0161ance je, \u017ee n\u011bjakou takovou zranitelnost obsahuje.<\/p>\n Velk\u00e9 softwarov\u00e9 firmy obvykle zam\u011bstn\u00e1vaj\u00ed cel\u00fd t\u00fdm lid\u00ed, kte\u0159\u00ed takov\u00e9 zranitelnosti hledaj\u00ed. Nav\u00edc nab\u00edzej\u00ed odm\u011bny (tzv. bug bounty) komukoli, kdo dok\u00e1\u017ee naj\u00edt a\u00a0zdokumentovat chybu, kter\u00e1 by mohla v\u00e9st k\u00a0obejit\u00ed autorizace nebo jin\u00e9mu nebezpe\u010dn\u00e9mu chov\u00e1n\u00ed.<\/p>\n Kdy\u017e ale n\u011bkdo\u00a0\u2013 \u0159ekn\u011bme nez\u00e1visl\u00e1 hackerka\u00a0\u2013 takovou zranitelnost objev\u00ed, m\u00e1 na v\u00fdb\u011br: M\u016f\u017ee ji ozn\u00e1mit firm\u011b, kter\u00e1 software vyv\u00edj\u00ed. Dostane pen\u00edze, zranitelnost bude v\u00a0dohledn\u00e9 dob\u011b zacelena, sv\u011bt se to\u010d\u00ed d\u00e1l. Anebo m\u016f\u017ee zranitelnost prodat na \u010dern\u00e9m trhu.<\/p>\n Koup\u00ed ji n\u011bjak\u00e1 hackersk\u00e1 skupina, nebo (nejsp\u00ed\u0161e skrze prost\u0159edn\u00edka) tajn\u00e9 slu\u017eby. Ty maj\u00ed velk\u00fd z\u00e1jem na tom, aby m\u011bly v\u00a0z\u00e1sob\u011b d\u016fle\u017eit\u00e9 a\u00a0dosud neodhalen\u00e9 zranitelnosti, kter\u00e9 pak mohou vyu\u017e\u00edt pro kybernetick\u00e9 \u00fatoky nebo p\u0159i prolamov\u00e1n\u00ed \u0161ifrov\u00e1n\u00ed, vzd\u00e1len\u00e9m odposlouch\u00e1v\u00e1n\u00ed a\u00a0tvorb\u011b \u201ezadn\u00edch vr\u00e1tek\u201c.<\/p>\n Nov\u00e1 kybernetick\u00e1 realita<\/p>\n S n\u00e1stupem um\u011bl\u00e9 inteligence schopn\u00e9 programovat<\/a> se samoz\u0159ejm\u011b zv\u00fd\u0161il po\u010det vyprodukovan\u00fdch \u0159\u00e1dk\u016f k\u00f3du a\u00a0s t\u00edm i\u00a0po\u010det t\u011bchto bezpe\u010dnostn\u00edch mezer. Do roku 2025\u00a0byla AI v\u00a0bezpe\u010dnostn\u00edch kruz\u00edch vn\u00edman\u00e1 p\u0159edev\u0161\u00edm jako nebezpe\u010dn\u00fd zdroj nov\u00fdch zranitelnost\u00ed. Lid\u00e9, kte\u0159\u00ed nikdy neprogramovali, najednou na web pou\u0161t\u011bli jednu \u201enavajbk\u00f3d\u011bnou\u201c aplikaci za druhou a\u00a0tyto aplikace \u010dasto obsahuj\u00ed \u0161kol\u00e1ck\u00e9 chyby<\/a>.<\/p>\n Koncem roku 2025\u00a0se ale za\u010daly AI modely zlep\u0161ovat natolik, \u017ee dok\u00e1zaly chyby nejen vytv\u00e1\u0159et, ale tak\u00e9 aktivn\u011b odhalovat. Schopnosti hledat \u201eaktivn\u00ed zranitelnosti\u201c v\u00a0k\u00f3du p\u0159itom rychle rostou. Doned\u00e1vna nejsiln\u011bj\u0161\u00ed model od Anthropic\u00a0\u2013 Opus 4.6\u00a0z\u00a0\u00fanora 2026\u00a0\u2013 byl schopen naj\u00edt zranitelnost<\/a>, kter\u00e1 se v\u00a0j\u00e1d\u0159e opera\u010dn\u00edho syst\u00e9mu Linux skr\u00fdvala 23\u00a0let.<\/p>\n \u201eKdy\u017e jsem stejnou \u00falohu zadal star\u0161\u00edm model\u016fm, tak nedok\u00e1zaly naj\u00edt skoro \u017e\u00e1dnou zranitelnost,\u201c uvedl<\/a> na ned\u00e1vn\u00e9 p\u0159edn\u00e1\u0161ce Nicholas Carlini, v\u00fdzkumn\u00edk spole\u010dnosti Anthropic. \u201eJe vid\u011bt, \u017ee se modely v\u00a0hled\u00e1n\u00ed bug\u016f v\u00fdrazn\u011b zlep\u0161ily. A\u00a0Opus 4.6\u00a0nen\u00ed posledn\u00ed model, kter\u00fd bude existovat.\u201c<\/p>\n V grafu po\u010dtu nalezen\u00fdch bug\u016f vid\u00edme starou zn\u00e1mou exponenci\u00e1lu\u2026<\/p>\n V dob\u011b, kdy Carlini prezentoval v\u00fdsledky modelu Opus 4.6, u\u017e nepochybn\u011b v\u011bd\u011bl o\u00a0nov\u00e9m modelu Mythos. Ten m\u00e1 b\u00fdt v\u00a0odhalov\u00e1n\u00ed zranitelnost\u00ed je\u0161t\u011b lep\u0161\u00ed. A\u00a0to o\u00a0tolik lep\u0161\u00ed, \u017ee by takov\u00fd model byl nebezpe\u010dn\u00fd, kdyby k\u00a0n\u011bmu m\u011bl p\u0159\u00edstup kdokoli.<\/p>\n Nov\u00fd model Mythos um\u00ed bezpe\u010dnostn\u00ed d\u00edry nejen naj\u00edt, ale tak\u00e9 jich vyu\u017e\u00edt k\u00a0tomu, aby z\u00a0nich vyrobil zranitelnost, kter\u00e9 m\u016f\u017ee n\u011bkdo zneu\u017e\u00edt, popsal\u00a0\u2013 nyn\u00ed u\u017e po ozn\u00e1men\u00ed Mythosu<\/a>\u00a0\u2013 Carlini: \u201eModel um\u00ed z\u0159et\u011bzit zranitelnosti. To znamen\u00e1, \u017ee model dok\u00e1\u017ee naj\u00edt n\u011bkolik zranitelnost\u00ed, kter\u00e9 by \u00fato\u010dn\u00edkovi samy o\u00a0sob\u011b nic moc neposkytly. Ale z\u0159et\u011bzen\u00edm t\u0159\u00ed, \u010dty\u0159, p\u011bti takov\u00fdch zranitelnost\u00ed dok\u00e1\u017ee Mythos vytvo\u0159it velmi sofistikovan\u00fd \u00fatok.\u201c<\/p>\n Jak si stoj\u00ed Mythos v\u00a0konkurenci ostatn\u00edch?<\/p>\n V porovn\u00e1n\u00ed s\u00a0dosud nejsiln\u011bj\u0161\u00edm modelem od Anthropic (Claude Opus 4.6\u00a0z\u00a0\u00fanora 2026) m\u00e1 Claude Mythos Preview lep\u0161\u00ed v\u00fdsledky v\u00a0mnoha ohledech. Dokumentace<\/a> zv\u00fdraznila nap\u0159\u00edklad tyto v\u00fdsledky test\u016f (kter\u00e9 z\u00a0d\u016fvodu nep\u0159\u00edstupnosti Mythosu nelze nez\u00e1visle ov\u011b\u0159it).<\/p>\n TestCo testuje?Opus 4.6Mythos PreviewSWE-bench ProProgramov\u00e1n\u00ed53,4 %77,8 %GPQA Diamond<\/p>\n Uva\u017eov\u00e1n\u00ed91,3 %94,6 %HLEPokro\u010dil\u00e9 uva\u017eov\u00e1n\u00ed53,1 %64,7 %CyberGymKyberbezpe\u010dnost0,670,83Firefox 147 exploitRe\u00e1ln\u00e9 prolomen\u00ed bezpe\u010dnosti0,8 %72,4 %\u00datok na firemn\u00ed s\u00ed\u0165Simulace re\u00e1ln\u00e9ho sc\u00e9n\u00e1\u0159enevy\u0159e\u0161ilvy\u0159e\u0161il<\/p>\n Pr\u00e1v\u011b schopnost nach\u00e1zet bezpe\u010dnost\u00ed d\u00edry a\u00a0aktivn\u011b je zneu\u017e\u00edt k\u00a0proniknut\u00ed do firemn\u00ed s\u00edt\u011b vedla firmu Anthropic k\u00a0tomu, aby nov\u00fd model dala k\u00a0dispozici jen omezen\u00e9mu mno\u017estv\u00ed partner\u016f. <\/p>\n \u201e\u010casov\u00e9 okno mezi odhalen\u00edm zranitelnosti (bezpe\u010dnostn\u00ed mezery v\u00a0software, pozn. red.) a\u00a0jej\u00edm zneu\u017eit\u00edm kybernetick\u00fdmi \u00fato\u010dn\u00edky se dramaticky zmen\u0161ilo,\u201c popisuje novou situaci Elia Zaitsev, technologick\u00fd \u0161\u00e9f CrowdStrike, tedy jedn\u00e9 z\u00a0vyvolen\u00fdch firem, kter\u00e1 m\u00e1 k\u00a0pokro\u010dil\u00e9mu Mythosu p\u0159\u00edstup. \u201eTo, co d\u0159\u00edve trvalo m\u011bs\u00edce, se dnes d\u00edky um\u011bl\u00e9 inteligenci odehraje b\u011bhem n\u011bkolika minut.\u201c<\/p>\n Mythos podle n\u011bj n\u00e1zorn\u011b ukazuje, jak\u00e9 mo\u017enosti dnes maj\u00ed obr\u00e1nci p\u0159i plo\u0161n\u00e9 automatizovan\u00e9 ochran\u011b softwaru. \u201eJe v\u0161ak nevyhnuteln\u00e9, \u017ee se \u00fato\u010dn\u00edci pokus\u00ed tyt\u00e9\u017e schopnosti zneu\u017e\u00edt ve sv\u016fj prosp\u011bch,\u201c varuje Zaitsev.<\/p>\n S t\u00edm souzn\u00ed i\u00a0z\u00e1stupce dal\u0161\u00ed firmy, kter\u00e1 byla do elitn\u00edho spolku vpu\u0161t\u011bna: \u201eSchopnosti um\u011bl\u00e9 inteligence od z\u00e1klad\u016f m\u011bn\u00ed nal\u00e9havost, s\u00a0n\u00ed\u017e mus\u00edme chr\u00e1nit kritickou infrastrukturu p\u0159ed kybernetick\u00fdmi hrozbami, a\u00a0nen\u00ed cesty zp\u011bt,\u201c uvedl Anthony Grieco, bezpe\u010dnostn\u00ed \u0161\u00e9f firmy Cisco. \u201eNa\u0161e pr\u00e1ce s\u00a0t\u011bmito modely uk\u00e1zala, \u017ee dok\u00e1\u017eeme odhalovat a\u00a0opravovat bezpe\u010dnostn\u00ed trhliny nap\u0159\u00ed\u010d hardwarem i\u00a0softwarem v\u00a0d\u0159\u00edve naprosto nemysliteln\u00e9m tempu a\u00a0rozsahu. Jde o\u00a0z\u00e1sadn\u00ed zlom a\u00a0jasn\u00fd sign\u00e1l, \u017ee star\u00e9 zp\u016fsoby procesu preventivn\u00edho zabezpe\u010dov\u00e1n\u00ed syst\u00e9m\u016f u\u017e zkr\u00e1tka nesta\u010d\u00ed.\u201c<\/p>\n U\u017e je to tady<\/p>\n \u201eTe\u010f m\u00e1me v\u0161ichni k\u00a0dispozici nejnov\u011bj\u0161\u00ed modely um\u011bl\u00e9 inteligence. Prakticky za pakatel, za p\u00e1r stovek m\u011bs\u00ed\u010dn\u011b. Nezapome\u0148me, \u017ee toto obdob\u00ed nebude trvat v\u011b\u010dn\u011b,\u201c poznamenal jsem ned\u00e1vno b\u011bhem jedn\u00e9 debaty o\u00a0AI.<\/p>\n P\u0159ipadal jsem si p\u0159i tom trochu jako nositel alobalov\u00e9 \u010depi\u010dky. V\u017edy\u0165 v\u0161echny velk\u00e9 i\u00a0mal\u00e9 firmy zat\u00edm\u00a0\u2013 minim\u00e1ln\u011b od roku 2023\u00a0\u2013 posp\u00edchaly s\u00a0ka\u017ed\u00fdm nov\u00fdm AI modelem rychle na trh. Kdo platil z\u00e1kladn\u00ed p\u0159edplatn\u00e9, m\u011bl v\u017edy p\u0159\u00edstup k\u00a0tomu nejlep\u0161\u00edmu v\u00a0oblasti generativn\u00ed AI.<\/p>\n Firmy nav\u00edc provoz sv\u00fdch pokro\u010dil\u00fdch syst\u00e9m\u016f dotovaly, proto\u017ee od u\u017eivatel\u016f na opl\u00e1tku z\u00edsk\u00e1valy informace o\u00a0tom, jak lze tyto n\u00e1stroje nasadit nebo jak je vylep\u0161it. Od za\u010d\u00e1tku se ale spekulovalo o\u00a0tom, \u017ee ve chv\u00edli, kdy budou dostate\u010dn\u011b siln\u00e9, tak je\u00a0\u2013 samoz\u0159ejm\u011b v\u00a0z\u00e1jmu bezpe\u010dnosti\u00a0\u2013 nebude mo\u017en\u00e9 pustit ven jen tak, pro v\u0161echny.<\/p>\n![](\"https:\/\/d39-a.sdn.cz\/d_39\/c_img_qA_A\/kcHp2YdDtBRc2UPz0GSWW6O\/59a0\/bug.png?fl=cro,0,0,688,413%7Cres,1200,,1%7Cwebp,75\")
Foto: Wikimedia<\/p>\n![](\"https:\/\/d39-a.sdn.cz\/d_39\/c_img_qA_A\/nCTUX1iX5CNPfpnlFGSWYIw\/8840\/vyvoj-ai-systemu.png?fl=cro,0,0,691,320%7Cres,1200,,1%7Cwebp,75\")
Foto: Nicholas Carlini\/YouTube<\/p>\n
![](\"https:\/\/d39-a.sdn.cz\/d_39\/c_img_p9_A\/kcHp2YdDtBM3KVE8kGGyzBB\/25a0\/ai.png?fl=cro,0,0,1920,1080%7Cres,320,,1%7Cwebp,75\")
<\/a><\/p>\n