OpenAI Ivula i-GPT-5.6-Cyber kubavikeli abagunyaziwe

Kwenzekeni

I-OpenAI inwebise uhlelo lwayo lwe-Daybreak ngo-Agasti 10 ngezigaba ezihlukene zokufinyelela zomsebenzi wokuzivikela kanye nocwaningo oluthuthukisiwe lwezokuphepha, okuhlanganisa nemodeli ekhethekile ye-GPT-5.6-Cyber.

I-Daybreak Blue inikeza abavikeli abagunyaziwe i-GPT-5.6 Sol enezihlungi ze-inthanethi zezinga lesistimu ezikhishiwe ukuze zisebenze njengokubuyekezwa kwekhodi evikelekile, ukuhlaziya uhlelo olungayilungele ikhompuyutha, impendulo yesigameko, ukutholwa kokuba sengozini, nokuqinisekiswa kwesiqephu. I-Daybreak Red yengeza i-GPT-5.6-Cyber ​​ukuze kuqinisekiswe ukuxhashazwa okugunyaziwe, ukuhlola ukungena, nolunye ucwaningo olusebenzisa kabili imodeli evamile engase ilunqabe.

I-OpenAI ithi i-GPT-5.6-Cyber ​​iqedele u-95% wezicelo ekuhlolweni kwayo kwangaphakathi kwe-Advanced Cybersecurity Completion Rate, uma kuqhathaniswa no-1.5% we-GPT-5.6 Sol ejwayelekile, u-2% we-Sol nge-Daybreak Blue, kanye no-57.3-1% we-ZXZAIU-Cyber. Le nkampani ithi ukuhlolwa kuhlanganisa izimo ezifana nokuthuthukiswa kwe-exploit-chain, i-bypass yokuqinisekisa, kanye nokwenyuka kwamalungelo; ayikakhiphi isethi yokuhlola.

Inkampani iphinde ibike ngokusebenzisa imodeli ukusiza ukuthola ubungozi be-V8 obungaziwa ngaphambilini obungaboshwa ndawonye. I-Google yabela eyokuqala i-CVE-2026-15903 futhi yayisipesheli ku-Chrome 150. Iseluleko sokukhishwa kwe-Google siqinisekisa ngokuzimele ukuthi iphutha lokuqina eliphezulu livumele ukufundwa nokubhala ngaphandle kwemingcele ku-V8, nakuba ingaluhloli uhlelo lwe-AI olusize ukuyithola.

I-OpenAI ichaza i-GPT-5.6-Cyber ​​njengemodeli eyakhelwe ku-GPT-5.6 Sol futhi eqeqeshelwe imisebenzi enobungozi obukhulu, esetshenziswa kabili esikhundleni sokusebenzisa okuhlaselayo okungakhawulelwe. I-Daybreak Red itholakala kuphela kubantu abagunyaziwe nezinhlangano ezenza umsebenzi ogunyaziwe, ngokuqinisekiswa komazisi, ukuphepha kwe-akhawunti, ukugadwa, imikhawulo yokusetshenziswa egunyaziwe, kanye nobufakazi obusemthethweni. Isimemezelo siphinde sithi ama-akhawunti e-Daybreak azodinga okhiye bokuphepha behadiwe kusukela ngoSepthemba 1, 2026.

Imininingwane yomthombo: OpenAI's GPT-5.6-Cyber and Daybreak announcement ↗

Kungani kubalulekile

Ukukhishwa kuhambisa amandla e-inthanethi ayingozi kakhulu emodeli yomngcele ukusuka kusilungiselelo esincane socwaningo kuya ohlelweni lokufinyelela olulawulwayo, ukuhlola ukuthi ingabe abavikeli bangawathola yini ngaphambi kokuthi amasu afanayo asabalale kubahlaseli.

Ukunciphisa ukwenqatshwa kuyisisekelo somkhiqizo: imodeli enqaba umsebenzi wokuxhaphaza ongokoqobo ingase iphephe kakhudlwana ukuze isetshenziswe okuvamile kodwa ingasebenziseki kangako emaqenjini okufanele akhiqize kabusha ubungozi ngaphambi kokuba abeke kuqala futhi akulungise. I-Daybreak ihlukanisa imisebenzi yokuzivikela ebanzi kusukela ekuthuthukisweni kokuxhashazwa okuyingozi kakhulu esikhundleni sokuveza inqubomgomo eyodwa kuwo wonke umsebenzisi.

Ukudalulwa kwe-V8 kuwubufakazi obuphathekayo bokuthi ucwaningo olusizwa yi-AI lokuba sengozini lungafinyelela isofthiwe esetshenzisiwe. Futhi kuncane kunesimangalo sokuzivikela okuzimele: I-OpenAI ithi abacwaningi bayo baqinisekise okutholakele nokudalulwa okudidiyelwe ne-Google, kuyilapho izeluleko zomphakathi ze-Chrome zirekhoda iphutha elixoshiwe esikhundleni sendima yemodeli enqubweni ephelele yocwaningo.

I-OpenAI futhi yandisa i-Daybreak ngokusebenzisa izixhumanisi zokuphepha nabahlinzeki bezobuchwepheshe. Ozakwethu abagunyaziwe bagcina ukufinyelela kumamodeli ayisisekelo futhi bawasebenzise ngaphakathi kokusebenzelana kwamakhasimende, ngezilawuli ezingabandakanya ukuqinisekiswa kobunikazi, izikophu ezichaziwe, ukuloga, ukugada, nokubuyekezwa komuntu.

Lowo mklamo uzakwethu uyashintsha ukuthi ubani okufanele anikeze imayela lokugcina lokwahlulela. I-OpenAI ithi ukufinyelela kumodeli engaphansi akudluliselwa ngokuqondile kukhasimende; uzakwethu uchaza ukuhlanganyela, ubuyekeze okutholiwe, futhi asebenzise ubuchwepheshe bakhe bokuphepha ngaphambi kwesenzo. Empeleni, ukusetshenziswa okuwusizo ngakho-ke kuncike emandleni kazakwethu okuhlukanisa iphutha elingaphinda likhiqizeke esiphakamisweni semodeli yokuqagela, aqhubeke nokuhlola ngaphakathi kwesikophu okuvunyelwene ngaso, futhi ahambise ukulungiswa okuqinisekisiwe ngenqubo yokunakekela isofthiwe yangempela.

Ongakubuka ngokulandelayo

Buka ikhadi lesistimu elithenjisiwe, ukukhiqizwa kabusha kokulinganisa okuzimele, ukudalulwa okuqediwe kobunye ubungozi obubikiwe, nobufakazi bokuthi izilawuli zokufinyelela zimelana nokusetshenziswa kabi.

Izibalo eziningi zokusebenza esimemezelweni zivela ekusetshenzisweni kwangaphakathi kwe-OpenAI kwe-Advanced Cybersecurity Completion Rate, i-ExploitGym, i-ExploitBench, nokuhlolwa kokubika ubungozi. Inkampani ithi i-GPT-5.6-Cyber ​​ayizange iphumelele zonke iziqhathaniso: GPT-5.6 Sol ikhiqize imibiko engcono ngokuhlola okukodwa futhi yaxazulula isilungiselelo esivamile se-300-turn ExploitBench ngempumelelo kakhudlwana.

I-OpenAI ilinganisela i-GPT-5.6-Cyber ​​at Phezulu, kodwa ngaphansi Okubalulekile, ngamakhono okuphepha ku-inthanethi ngaphansi kohlaka lwayo Lokulungiselela. Ithi ikhadi lesistimu eligcwele lizofika kamuva. Kuze kube yilapho izindlela nemiphumela ishicilelwa ngemininingwane eyanele ukuze iphinde ikhiqizwe, isimemezelo sisekela izimangalo zenkampani okubaluliwe kunesilinganiso esizimele samandla okuhlasela komhlaba wangempela.

Icala lokuphepha lincike ekusebenzeni okuningi njengokuqeqeshwa. Ama-akhawunti e-Daybreak ngamunye kufanele asebenzise okhiye bokuqinisekisa ubunikazi bezingxenyekazi zekhompuyutha kusukela ngo-September 1, futhi i-OpenAI incoma izindawo ezingazodwa, izimvume ezinobubanzi, izenzo ze-ejenti ezigadiwe, nokugadwa komuntu. Ukubika okuwusizo kokulandelela kufanele kubonise ukuhoxiswa kokufinyelela, ukuhlukumeza okutholiwe, izikhathi zokuveza ulwazi, imiphumela yokuchibiyela, kanye nokuhle okungamanga njengoba uhlelo lukhula.

Ukuqhathaniswa kwesilinganiso sokuqedwa kwesihloko esikhulu nakho kune-caveat yokulinganisa ebalulekile. I-OpenAI ithi yonke imodeli yaqhutshwa ngezinga eliphezulu lokucabanga elitholakala esidlangalaleni, futhi iphawula ukuthi i-GPT-5.6-Cyber ​​ivamise ukusebenzisa ibhajethi yokucabanga enkulu kanye namathokheni amaningi kune-GPT-5.6 Sol. Ngakho-ke izinga eliphezulu lokuqeda lingabonisa kokubili ukuqeqeshwa okukhethekile kanye nenani elihlukile lokubala. Ukuhlola okuzimele kufanele kubike amabhajethi afanayo, ukwakhiwa komsebenzi, imibandela yokwenqaba, kanye nezinga lezimpendulo ezingaphephile noma ezingasebenziseki kunokuba kuphela ukuthi impendulo yakhiqizwa yini.

I-akhawunti ekhulile yohlelo izodinga ukulandela wonke umjikelezo wempilo wokuba sengozini. Lokho kusho ukubika ukuthi kukangakhi umthofu okhiqizwe imodeli usinda ekukhiqizeni kabusha komuntu, ukuthi kungakanani okutholiwe okuyimpinda noma ubukhali obuphansi, ukuthi abathengisi bathola ngokushesha kangakanani imibiko engenzeka, kanye nokuthi amapeshi asetshenziswa ngaphambi kokudalulwa komphakathi. Kufanele futhi ichaze ukuthi kwenzekani lapho ikhasimende likazakwethu licela imodeli ukuthi yeqe umngcele wokuhlola, lapho umenzeli ehlangabezana nemininingwane yokukhiqiza, noma lapho ukuxhaphaza okuhlongozwayo kuyingozi kakhulu ukuthi kungagunyazwa ngokuqondile. Lezo zilawuli zinquma ukuthi ingabe izilinganiso zokwenqaba eziphansi ziyawakha yini inani elilinganisekayo lokuvikela esikhundleni sokumane andise ukunikezwa kwekhodi yokusetshenziswa okukabili.

Umphakathi kufanele futhi ukwazi ukuhlukanisa ikhono lemodeli nokugunyazwa yikhasimende. Imithetho yokufinyelela ye-OpenAI ibeka iziqinisekiso ezingokomthetho, ukuhlolwa kobunikazi, ukugada, nomsebenzi osuselwe kumodeli, kodwa lezo zilawuli ziyizicelo ezizohlolwa ngokuhamba kwesikhathi. Ukubikwa kwesigameko kufanele kusho ukuthi isicelo sivinjiwe yini, sikhushulelwe ukuthi sibuyekezwe, noma sivunyelwe ngaphansi kokuhlanganyela okubhaliwe, ngaphandle kokuveza imininingwane yokuxhaphaza engenza isistimu enganyathelisiwe ibe lula ukuhlasela.

Imihlahlandlela ehlobene nemibuzo