{"id":167,"date":"2026-02-05T20:33:34","date_gmt":"2026-02-05T19:33:34","guid":{"rendered":"https:\/\/helloblog.io\/cs\/gpt-5-3-codex-codex-agent-pro-praci-na-pocitaci\/"},"modified":"2026-02-05T20:33:34","modified_gmt":"2026-02-05T19:33:34","slug":"gpt-5-3-codex-codex-agent-pro-praci-na-pocitaci","status":"publish","type":"post","link":"https:\/\/helloblog.io\/cs\/gpt-5-3-codex-codex-agent-pro-praci-na-pocitaci\/","title":{"rendered":"GPT\u20115.3\u2011Codex: Codex u\u017e nen\u00ed jen \u201ecoding agent\u201c, ale prakticky kolega pro pr\u00e1ci na po\u010d\u00edta\u010di"},"content":{"rendered":"\n

OpenAI 5. \u00fanora 2026 p\u0159edstavilo GPT\u20115.3\u2011Codex \u2013 nov\u00fd model pro Codex, kter\u00fd m\u00e1 ambici pokr\u00fdt prakticky cel\u00e9 spektrum profesion\u00e1ln\u00ed pr\u00e1ce na po\u010d\u00edta\u010di. Nejde jen o dal\u0161\u00ed iteraci \u201emodelu na k\u00f3d\u201c: podle ozn\u00e1men\u00ed kombinuje frontier v\u00fdkon GPT\u20115.2\u2011Codex v programov\u00e1n\u00ed s rozumov\u00fdmi a znalostn\u00edmi schopnostmi GPT\u20115.2 do jednoho modelu, a z\u00e1rove\u0148 b\u011b\u017e\u00ed o 25 % rychleji<\/strong>. V praxi to znamen\u00e1 del\u0161\u00ed b\u011bhy \u00faloh, v\u00edc pr\u00e1ce s n\u00e1stroji (tool use) a v\u00edc komplexn\u00edho vykon\u00e1v\u00e1n\u00ed krok\u016f bez toho, aby se ztr\u00e1cel kontext.<\/p>\n\n\n\n

Zaj\u00edmav\u00fd detail: GPT\u20115.3\u2011Codex je podle OpenAI prvn\u00ed model, kter\u00fd byl \u201einstrument\u00e1ln\u00ed\u201c p\u0159i vytv\u00e1\u0159en\u00ed sebe sama. T\u00fdm Codexu pou\u017eil ran\u00e9 verze k lad\u011bn\u00ed vlastn\u00edho tr\u00e9ninku, k \u0159\u00edzen\u00ed nasazen\u00ed i k diagnostice test\u016f a evaluac\u00ed. Jin\u00fdmi slovy: model zrychloval v\u00fdvoj vlastn\u00edho n\u00e1stupce.<\/p>\n\n\n\n

A je\u0161t\u011b jedna posunov\u00e1 v\u011bta, kter\u00e1 dob\u0159e vystihuje sm\u011br: Codex se t\u00edmhle krokem posouv\u00e1 z agenta, kter\u00fd p\u00ed\u0161e a kontroluje k\u00f3d, na agenta, kter\u00fd zvl\u00e1dne t\u00e9m\u011b\u0159 cokoli, co v\u00fdvoj\u00e1\u0159i a dal\u0161\u00ed profesion\u00e1lov\u00e9 b\u011b\u017en\u011b d\u011blaj\u00ed na po\u010d\u00edta\u010di.<\/p>\n\n\n\n

Frontier agentn\u00ed schopnosti: co se vlastn\u011b zlep\u0161ilo<\/h2>\n\n\n\n

OpenAI v ozn\u00e1men\u00ed stav\u00ed GPT\u20115.3\u2011Codex jako nov\u00fd \u201eindustry high\u201c na benchmarkov\u00fdch sad\u00e1ch SWE\u2011Bench Pro<\/strong> a Terminal\u2011Bench<\/strong> a z\u00e1rove\u0148 zmi\u0148uje siln\u00fd v\u00fdkon na OSWorld<\/strong> a GDPval<\/strong>. Tyhle \u010dty\u0159i benchmarky pou\u017e\u00edvaj\u00ed intern\u011b pro m\u011b\u0159en\u00ed programov\u00e1n\u00ed, agentn\u00edho chov\u00e1n\u00ed a \u201ereal\u2011world\u201c schopnost\u00ed.<\/p>\n\n\n\n

Programov\u00e1n\u00ed: SWE\u2011Bench Pro a Terminal\u2011Bench 2.0<\/h3>\n\n\n\n

V \u010dist\u011b softwarov\u00e9m in\u017een\u00fdrstv\u00ed se GPT\u20115.3\u2011Codex podle OpenAI dost\u00e1v\u00e1 na state\u2011of\u2011the\u2011art na SWE\u2011Bench Pro<\/strong>. Tohle je podstatn\u00e9 i kv\u016fli tomu, co se testuje: zat\u00edmco SWE\u2011bench Verified je omezen\u00fd na Python, SWE\u2011Bench Pro pokr\u00fdv\u00e1 \u010dty\u0159i jazyky<\/strong> a m\u00e1 b\u00fdt odoln\u011bj\u0161\u00ed v\u016f\u010di kontaminaci (contamination\u2011resistant), n\u00e1ro\u010dn\u011bj\u0161\u00ed, pest\u0159ej\u0161\u00ed a bl\u00ed\u017e realit\u011b v pr\u016fmyslu.<\/p>\n\n\n\n

Sou\u010dasn\u011b model v\u00fdrazn\u011b p\u0159ekon\u00e1v\u00e1 p\u0159edchoz\u00ed state\u2011of\u2011the\u2011art na Terminal\u2011Bench 2.0<\/strong>, kter\u00fd c\u00edl\u00ed na \u201etermin\u00e1lov\u00e9\u201c dovednosti, kter\u00e9 agent typu Codex pot\u0159ebuje (typicky pr\u00e1ce s p\u0159\u00edkazy, iterativn\u00ed diagnostika, skriptov\u00e1n\u00ed a podobn\u011b). OpenAI nav\u00edc vypichuje, \u017ee GPT\u20115.3\u2011Codex tohle zvl\u00e1d\u00e1 s men\u0161\u00edm po\u010dtem token\u016f<\/strong> ne\u017e d\u0159\u00edv\u011bj\u0161\u00ed modely \u2013 co\u017e v praxi znamen\u00e1 lep\u0161\u00ed vyu\u017eit\u00ed kontextu a v\u00edc prostoru na u\u017eite\u010dnou pr\u00e1ci.<\/p>\n\n\n\n

Webov\u00fd v\u00fdvoj: dlouh\u00e9 b\u011bhy, estetika a \u201ecompaction\u201c<\/h3>\n\n\n\n

Vedle benchmark\u016f se OpenAI sna\u017e\u00ed uk\u00e1zat, \u017ee model um\u00ed d\u011blat v\u00fdrazn\u011b komplexn\u011bj\u0161\u00ed webov\u00e9 v\u011bci \u201eod nuly\u201c a vydr\u017eet na nich pracovat dlouho. Kombinace frontier k\u00f3dov\u00e1n\u00ed, zlep\u0161en\u00ed v estetice a tzv. compaction<\/strong> (v kontextu agent\u016f typicky zhu\u0161\u0165ov\u00e1n\u00ed\/komprimace pr\u016fb\u011b\u017en\u00e9ho kontextu tak, aby se dalo pokra\u010dovat ve velk\u00fdch b\u011bz\u00edch) m\u00e1 v\u00e9st k tomu, \u017ee model zvl\u00e1dne stav\u011bt funk\u010dn\u00ed hry a aplikace p\u0159es dny iterac\u00ed.<\/p>\n\n\n\n

Konkr\u00e9tn\u011b OpenAI otestovalo web dev i dlouhotrvaj\u00edc\u00ed agentn\u00ed chov\u00e1n\u00ed tak, \u017ee po\u017e\u00e1dalo GPT\u20115.3\u2011Codex o vytvo\u0159en\u00ed dvou her:<\/p>\n\n\n\n

    \n\n
  • druhou verzi z\u00e1vodn\u00ed hry z p\u0159edchoz\u00edho uveden\u00ed Codex appu (viz odkaz n\u00ed\u017ee),<\/li>\n\n\n
  • a pot\u00e1p\u011b\u010dskou hru zalo\u017eenou na pr\u016fzkumu \u00fates\u016f a sb\u011bratelsk\u00e9m \u201efish codex\u201c s managementem kysl\u00edku, tlaku a hazard\u016f.<\/li>\n\n<\/ul>\n\n\n\n

    Test prob\u00edhal s pou\u017eit\u00edm skillu \u201edevelop web game\u201c a s p\u0159edvybran\u00fdmi, obecn\u00fdmi follow\u2011up prompty typu \u201efix the bug\u201c nebo \u201eimprove the game\u201c. Model pak m\u011bl iterovat autonomn\u011b \u201eover millions of tokens\u201c.<\/p>\n\n\n\n

    OpenAI zve\u0159ejnilo i hrateln\u00e9 v\u00fdsledky:<\/p>\n\n\n\n

      \n\n
    • Z\u00e1vodn\u00ed hra s v\u00edce jezdci, osmi mapami a itemy na mezern\u00edk: https:\/\/cdn.openai.com\/gpt-examples\/7fc9a6cb-887c-4db6-98ff-df3fd1612c78\/racing_v2.html<\/li>\n\n\n
    • Pot\u00e1p\u011b\u010dsk\u00e1 hra s \u00fatesy, sb\u011brem, kysl\u00edkem\/tlakem a hazardy: https:\/\/cdn.openai.com\/gpt-examples\/7fc9a6cb-887c-4db6-98ff-df3fd1612c78\/diving_game.html<\/li>\n\n<\/ul>\n\n\n\n

      V b\u011b\u017en\u011bj\u0161\u00edm webu OpenAI tvrd\u00ed, \u017ee GPT\u20115.3\u2011Codex l\u00edp ch\u00e1pe z\u00e1m\u011br u \u201eday\u2011to\u2011day\u201c web\u016f ne\u017e GPT\u20115.2\u2011Codex: jednoduch\u00e9 nebo nedostate\u010dn\u011b specifikovan\u00e9 prompty pr\u00fd \u010dast\u011bji kon\u010d\u00ed str\u00e1nkou s rozumn\u00fdmi defaulty a v\u011bt\u0161\u00ed funk\u010dnost\u00ed, tedy lep\u0161\u00edm \u201epl\u00e1tnem\u201c pro dal\u0161\u00ed \u00fapravy.<\/p>\n\n\n\n

      Jako ilustraci uv\u00e1d\u00ed srovn\u00e1n\u00ed dvou landing pages (stejn\u00fd prompt pro GPT\u20115.3\u2011Codex i GPT\u20115.2\u2011Codex). GPT\u20115.3\u2011Codex podle popisu automaticky:<\/p>\n\n\n\n

        \n\n
      • zobrazil ro\u010dn\u00ed pl\u00e1n jako zlevn\u011bnou m\u011bs\u00ed\u010dn\u00ed cenu (m\u00edsto prost\u00e9ho p\u0159epo\u010dtu ro\u010dn\u00ed sumy), tak\u017ee sleva p\u016fsob\u00ed z\u00e1m\u011brn\u011bji a \u010diteln\u011bji,<\/li>\n\n\n
      • vytvo\u0159il automaticky p\u0159ech\u00e1zej\u00edc\u00ed carousel s testimonials se t\u0159emi odli\u0161n\u00fdmi citacemi (ne jen jednou), tak\u017ee str\u00e1nka p\u016fsob\u00ed kompletji a bl\u00ed\u017e produkci u\u017e v z\u00e1kladu.<\/li>\n\n<\/ul>\n\n\n\n

        Mimo k\u00f3d: cel\u00fd software lifecycle i \u201eknowledge work\u201c<\/h3>\n\n\n\n

        Siln\u00e1 \u010d\u00e1st ozn\u00e1men\u00ed je p\u0159esun d\u016frazu z pouh\u00e9ho generov\u00e1n\u00ed k\u00f3du na podporu pr\u00e1ce nap\u0159\u00ed\u010d software lifecycle: debugging, deployment, monitoring, psan\u00ed PRD (Product Requirements Document), editace text\u016f, user research, testy, metriky a dal\u0161\u00ed. A z\u00e1rove\u0148 p\u0159izn\u00e1n\u00ed, \u017ee spousta rol\u00ed (engineer, designer, PM, data scientist) tr\u00e1v\u00ed v\u00fdrazn\u00fd \u010das v\u00fdstupy, kter\u00e9 nejsou k\u00f3d \u2013 prezentace, tabulky, dokumentace, anal\u00fdzy.<\/p>\n\n\n\n

        K tomu OpenAI odkazuje na GDPval<\/strong> (eval z roku 2025), kter\u00fd m\u011b\u0159\u00ed v\u00fdkon na dob\u0159e specifikovan\u00fdch \u00faloh\u00e1ch znalostn\u00ed pr\u00e1ce nap\u0159\u00ed\u010d 44 profesemi<\/strong>. GPT\u20115.3\u2011Codex m\u00e1 na GDPval vykazovat siln\u00fd v\u00fdkon a podle ozn\u00e1men\u00ed \u201ematching GPT\u20115.2\u201c.<\/p>\n\n\n\n

        V \u010dl\u00e1nku jsou uk\u00e1zky typ\u016f v\u00fdstup\u016f, kter\u00e9 agent v r\u00e1mci t\u011bchto \u00faloh vytv\u00e1\u0159el (nap\u0159. slides s finan\u010dn\u00edm poradenstv\u00edm, \u0161kol\u00edc\u00ed dokument pro retail, NPV anal\u00fdzu v tabulce, fashion prezentaci do PDF). Jeden z uveden\u00fdch p\u0159\u00edklad\u016f obsahuje i cel\u00fd kontext zad\u00e1n\u00ed: agent v roli finan\u010dn\u00edho poradce p\u0159ipravuje 10slidovou prezentaci, pro\u010d jako fiduciary doporu\u010dit klient\u016fm nevolit p\u0159evod CD (certificates of deposits) do variable annuities, s explicitn\u00edmi srovn\u00e1n\u00edmi a oporou ve zdroj\u00edch NAIC a FINRA.<\/p>\n\n\n\n

        \"Uk\u00e1zka
        Uk\u00e1zkov\u00fd v\u00fdstup (GDPval) publikovan\u00fd v ozn\u00e1men\u00ed OpenAI. \u2014 Forr\u00e1s: OpenAI<\/em><\/figcaption><\/figure>\n\n\n\n

        OpenAI p\u0159ipom\u00edn\u00e1, \u017ee jednotliv\u00e9 \u00falohy v GDPval navrhuj\u00ed zku\u0161en\u00ed profesion\u00e1lov\u00e9 a maj\u00ed odpov\u00eddat re\u00e1ln\u00e9 pr\u00e1ci v dan\u00e9 profesi.<\/p>\n\n\n\n

        Dal\u0161\u00ed d\u016fle\u017eit\u00fd benchmark je OSWorld<\/strong>: agent m\u00e1 plnit produktivn\u00ed \u00fakoly ve vizu\u00e1ln\u00edm desktopov\u00e9m prost\u0159ed\u00ed, a model pou\u017e\u00edv\u00e1 vision. OpenAI tvrd\u00ed, \u017ee GPT\u20115.3\u2011Codex zde ukazuje v\u00fdrazn\u011b lep\u0161\u00ed schopnosti pr\u00e1ce s po\u010d\u00edta\u010dem ne\u017e p\u0159edchoz\u00ed GPT modely. V OSWorld\u2011Verified se v \u010dl\u00e1nku zmi\u0148uje, \u017ee lid\u00e9 dosahuj\u00ed p\u0159ibli\u017en\u011b ~72 %.<\/p>\n\n\n\n

        Celkov\u00fd r\u00e1mec, kter\u00fd OpenAI stav\u00ed, je jasn\u00fd: kombinace v\u00fdsledk\u016f v k\u00f3dov\u00e1n\u00ed, frontendu a \u201ecomputer\u2011use\u201c \u00faloh\u00e1ch nazna\u010duje posun sm\u011brem k jednomu obecn\u011bj\u0161\u00edmu agentovi, kter\u00fd um\u00ed p\u0159em\u00fd\u0161let, stav\u011bt i vykon\u00e1vat kroky nap\u0159\u00ed\u010d re\u00e1lnou technickou prac\u00ed.<\/p>\n\n\n\n

        Interaktivn\u00ed spolupr\u00e1ce: m\u00e9n\u011b \u010dek\u00e1n\u00ed na fin\u00e1le, v\u00edc \u0159\u00edzen\u00ed v pr\u016fb\u011bhu<\/h2>\n\n\n\n

        S rostouc\u00edmi schopnostmi agent\u016f se podle OpenAI posouv\u00e1 bottleneck od \u201eco agent um\u00ed\u201c k tomu, jak snadno se d\u00e1 \u0159\u00eddit, zad\u00e1vat a dohl\u00ed\u017eet na v\u00edce agent\u016f paraleln\u011b. V tomhle sm\u011bru m\u00e1 Codex app zjednodu\u0161ovat management agent\u016f a s GPT\u20115.3\u2011Codex se m\u00e1 st\u00e1t v\u00edc interaktivn\u00ed.<\/p>\n\n\n\n

        Praktick\u00fd rozd\u00edl je v pr\u016fb\u011b\u017en\u00fdch updatech: m\u00edsto \u010dek\u00e1n\u00ed na fin\u00e1ln\u00ed v\u00fdstup m\u00e1 Codex \u010dast\u011bji reportovat kl\u00ed\u010dov\u00e1 rozhodnut\u00ed a progres. V re\u00e1ln\u00e9m \u010dase se pak d\u00e1 dopt\u00e1vat, diskutovat p\u0159\u00edstupy a \u201esteerovat\u201c sm\u011brem k \u0159e\u0161en\u00ed. Model m\u00e1 pr\u016fb\u011b\u017en\u011b vysv\u011btlovat, co d\u011bl\u00e1, reagovat na zp\u011btnou vazbu a dr\u017eet u\u017eivatele v obraze od za\u010d\u00e1tku do konce.<\/p>\n\n\n\n

        Pokud chce\u0161 steering zapnout, v Codex appu je to podle ozn\u00e1men\u00ed v: Settings > General > Follow-up behavior<\/strong>.<\/p>\n\n\n\n

        Jak OpenAI pou\u017eilo Codex p\u0159i tr\u00e9ninku a nasazen\u00ed GPT\u20115.3\u2011Codex<\/h2>\n\n\n\n

        Z technick\u00e9ho pohledu je na ozn\u00e1men\u00ed nejzaj\u00edmav\u011bj\u0161\u00ed to, jak otev\u0159en\u011b popisuje, \u017ee Codex se stal intern\u00edm akceler\u00e1torem pro v\u00fdzkum i engineering. OpenAI p\u00ed\u0161e, \u017ee posledn\u00ed rychl\u00e9 zlep\u0161ov\u00e1n\u00ed Codexu stoj\u00ed na v\u00fdzkumn\u00fdch projektech nap\u0159\u00ed\u010d firmou (v horizontu m\u011bs\u00edc\u016f a\u017e let) \u2013 a z\u00e1rove\u0148 tyhle projekty samotn\u00e9 Codex urychluje. N\u011bkte\u0159\u00ed v\u00fdzkumn\u00edci a in\u017een\u00fd\u0159i to pr\u00fd popisuj\u00ed tak, \u017ee jejich pr\u00e1ce je dnes fundament\u00e1ln\u011b jin\u00e1 ne\u017e p\u0159ed dv\u011bma m\u011bs\u00edci.<\/p>\n\n\n\n

        U\u017e ran\u00e9 verze GPT\u20115.3\u2011Codex m\u011bly podle OpenAI natolik v\u00fdjime\u010dn\u00e9 schopnosti, \u017ee t\u00fdm s nimi dok\u00e1zal zlep\u0161ovat tr\u00e9nink a podporovat nasazen\u00ed pozd\u011bj\u0161\u00edch verz\u00ed. Zm\u00edn\u011bn\u00e9 p\u0159\u00edklady pou\u017eit\u00ed jsou hodn\u011b konkr\u00e9tn\u00ed:<\/p>\n\n\n\n

          \n\n
        • V\u00fdzkumn\u00fd t\u00fdm pou\u017eil Codex k monitoringu a lad\u011bn\u00ed tr\u00e9novac\u00edho b\u011bhu pro tento release.<\/li>\n\n\n
        • Codex ne\u0159e\u0161il jen infrastrukturu: pom\u00e1hal sledovat vzorce v pr\u016fb\u011bhu tr\u00e9ninku, d\u011blal hlubokou anal\u00fdzu kvality interakc\u00ed, navrhoval opravy a stav\u011bl bohat\u00e9 aplikace pro lidi, aby p\u0159esn\u011b vid\u011bli, jak se chov\u00e1n\u00ed modelu li\u0161\u00ed od p\u0159edchoz\u00edch verz\u00ed.<\/li>\n\n\n
        • Engineering t\u00fdm s Codexem optimalizoval a adaptoval \u201eharness\u201c (testovac\u00ed\/integra\u010dn\u00ed r\u00e1mec) pro GPT\u20115.3\u2011Codex.<\/li>\n\n\n
        • Kdy\u017e se za\u010daly objevovat zvl\u00e1\u0161tn\u00ed edge\u2011casy dopadaj\u00edc\u00ed na u\u017eivatele, Codex pomohl identifikovat bugy v renderov\u00e1n\u00ed kontextu a naj\u00edt root cause n\u00edzk\u00fdch cache hit rates.<\/li>\n\n\n
        • B\u011bhem launch f\u00e1ze m\u00e1 GPT\u20115.3\u2011Codex d\u00e1l pom\u00e1hat dynamicky \u0161k\u00e1lovat GPU clustery kv\u016fli traffic \u0161pi\u010dk\u00e1m a dr\u017eet stabiln\u00ed latenci.<\/li>\n\n\n
        • V alpha testov\u00e1n\u00ed cht\u011bl jeden v\u00fdzkumn\u00edk odhadnout, kolik pr\u00e1ce nav\u00edc model ud\u011bl\u00e1 \u201eper turn\u201c a jak se li\u0161\u00ed produktivita. GPT\u20115.3\u2011Codex navrhl n\u011bkolik jednoduch\u00fdch regex klasifik\u00e1tor\u016f (frekvence up\u0159es\u0148uj\u00edc\u00edch dotaz\u016f, pozitivn\u00ed\/negativn\u00ed reakce u\u017eivatele, progres na \u00faloze) a \u0161k\u00e1lovateln\u011b je spustil nad session logy a vytvo\u0159il report se z\u00e1v\u011bry.<\/li>\n\n\n
        • Pozorov\u00e1n\u00ed z alpha: lid\u00e9 byli spokojen\u011bj\u0161\u00ed, proto\u017ee agent l\u00e9pe ch\u00e1pal z\u00e1m\u011br a ud\u011blal v\u00edc pr\u00e1ce na tah s m\u00e9n\u011b up\u0159es\u0148uj\u00edc\u00edmi ot\u00e1zkami.<\/li>\n\n\n
        • Data z alpha test\u016f m\u011bla kv\u016fli odli\u0161nosti modelu mnoho nezvykl\u00fdch a kontraintuitivn\u00edch v\u00fdsledk\u016f. Data scientist s pomoc\u00ed GPT\u20115.3\u2011Codex postavil nov\u00e9 datov\u00e9 pipeline a vizualizoval v\u00fdsledky bohat\u011bji, ne\u017e umo\u017e\u0148ovaly standardn\u00ed dashboardy.<\/li>\n\n\n
        • V\u00fdsledky pak spole\u010dn\u011b s Codexem ko\u2011analyzovali; Codex shrnul kl\u00ed\u010dov\u00e9 insighty p\u0159es tis\u00edce datov\u00fdch bod\u016f za m\u00e9n\u011b ne\u017e t\u0159i minuty.<\/li>\n\n<\/ul>\n\n\n\n

          OpenAI to uzav\u00edr\u00e1 t\u00edm, \u017ee jednotliv\u00e9 p\u0159\u00edklady jsou zaj\u00edmav\u00e9 samy o sob\u011b, ale dohromady pr\u00fd vedly k v\u00fdrazn\u00e9mu urychlen\u00ed pr\u00e1ce v\u00fdzkumu, engineeringu i produktov\u00fdch t\u00fdm\u016f.<\/p>\n\n\n\n

          Kyberbezpe\u010dnost: \u201eHigh capability\u201c klasifikace a siln\u011bj\u0161\u00ed ochrany<\/h2>\n\n\n\n

          V posledn\u00edch m\u011bs\u00edc\u00edch OpenAI pozoruje podle ozn\u00e1men\u00ed v\u00fdrazn\u00e9 zlep\u0161en\u00ed v\u00fdkonu na \u00faloh\u00e1ch kyberbezpe\u010dnosti \u2013 pro v\u00fdvoj\u00e1\u0159e i security profesion\u00e1ly. Paraleln\u011b firma p\u0159ipravovala pos\u00edlen\u00e9 ochrany (cyber safeguards) pro defenzivn\u00ed pou\u017eit\u00ed a celkovou odolnost ekosyst\u00e9mu.<\/p>\n\n\n\n

          GPT\u20115.3\u2011Codex je podle OpenAI prvn\u00ed model klasifikovan\u00fd pro kyberbezpe\u010dnostn\u00ed \u00falohy jako \u201eHigh capability\u201c<\/strong> v r\u00e1mci jejich Preparedness Framework<\/strong> a z\u00e1rove\u0148 prvn\u00ed model, kter\u00fd byl p\u0159\u00edmo tr\u00e9novan\u00fd na identifikaci softwarov\u00fdch zranitelnost\u00ed. OpenAI z\u00e1rove\u0148 \u0159\u00edk\u00e1, \u017ee nem\u00e1 definitivn\u00ed d\u016fkazy, \u017ee by model dok\u00e1zal automatizovat kyber\u00fatoky end\u2011to\u2011end \u2013 p\u0159esto vol\u00ed preventivn\u00ed (precautionary) p\u0159\u00edstup a nasazuje dosud nejkomplexn\u011bj\u0161\u00ed cyber security safety stack.<\/p>\n\n\n\n

          Konkr\u00e9tn\u011b vyjmenovan\u00e9 mitigace zahrnuj\u00ed:<\/p>\n\n\n\n