GPT-6 Cyber Uyasebenza

I-GPT-6 Cyber iyimodeli ye-OpenAI ekhethekile kwezokuphepha kwamasistimu e-computer, esekelwe ku-GPT-6 Astra futhi yaqeqeshwa ngokufunda okuqiniswayo kwemisebenzi yezokuphepha. Ngaphansi kohlaka lwe-Preparedness ithole isilinganiso esiphezulu se-Critical kwezokuphepha kwamasistimu e-computer, futhi itholakala kuphela kumakhasimende e-Daybreak Red agunyaziwe ngemva kokubuyekezwa. I-GPT-6 Cyber ingathola futhi isebenzise ubuthakathaka ngokuzimela: ikhiqize ama-exploit angu-948 okuthola amalungelo aphezulu ku-Linux kernel, yathola ubuthakathaka obungaziwa obuphezulu obungu-81 ku-Firefox, futhi yaqeda u-95% ku-RemoteExploitBench yokuhlola ukuhlasela okukude. Ngasikhathi sinye, i-GPT-6 Cyber yaqeqeshwa ngokukhethekile ukuze ihambisane nemithetho; ekuhlolweni kwe-honeypot okungaphandle kobubanzi yagcina imingcele yokugunyazwa ngo-99.93%, isekelwa ukuqapha kwabantu, uhlu olumhlophe lwama-IP, ukuvikelwa kwengqalasizinda ebalulekile kanye nezinye izithiyo eziningi.

Ukuthola ubuthakathaka: Ihlola ngokuzimela ubuthakathaka obungaziwa kumakhodi amakhulu afana ne-Linux kernel ne-Firefox, ikhiqize ubufakazi bokuhlola obungaphindwa.
Ukuthuthukisa ukuxhaphaza ubuthakathaka: Ibhalela ubuthakathaka obuqinisekisiwe izinhlelo zokuhlasela ezisebenzayo, futhi isikhiqize ama-exploit angu-948 okuthola amalungelo endaweni ku-Linux.
Ukuthuthukisa ama-patch: Ikhiqiza ama-patch aphakanyisiwe obuthakathaka obutholakele, isize ngokuqondile abagcini bamaphrojekthi ekulungiseni.
Ukuhlolwa kwamakhodi: Isiza amaqembu ezokuphepha ukuhlola amakhodi amakhulu, liqinisekise ubucayi bobuthakathaka nokuthi ukulungiswa kuphumelele yini.
Ukuhlolwa kokungena okugunyaziwe: Ngaphakathi kwemingcele egunyaziwe yenza uchungechunge oluphelele lokuhlasela ezindaweni ezikude ezifana ne-GitLab, PostgreSQL ne-Jenkins; iqede u-95% ku-RemoteExploitBench.
Ukuhlaziya ngokuhlehlisa: Yenza i-reverse engineering yezinhlelo ze-binary, yafinyelela ku-92.8% ku-SRE-Bench.
Ukuhlolwa kokungena ku-Web: Ihlola ubuthakathaka bezinhlelo ze-Web ngaphandle kokuba nekhodi yomthombo, ngezinga lokubuyisa elingu-40.3%.
Ukuzihlola kwengqalasizinda: Yasetshenziswa ekuhlaseleni indawo yokuqeqesha ye-OpenAI uqobo, yaphunyuka ku-microVM ngochungechunge lobuthakathaka, yathola futhi yaqinisa izindawo ezibuthaka ezinhlelweni zayo.

Isisekelo nokuqeqeshwa okuqondiswe ngemva kokuqeqeshwa: I-GPT-6 Cyber isekelwe ku-GPT-6 Astra. I-OpenAI yasebenzisa ukufunda okuqiniswayo ukwenza ukuqeqeshwa okukhethekile kwemisebenzi yezokuphepha, lapho imodeli iklonyeliswa ngokuphinda idale ukuphahlazeka, ithuthukise i-exploit esebenzayo, futhi ithole ubuthakathaka kukhodi; iphinde iklonyeliswe ngokugcina imingcele yokugunyazwa, ukungadluli imingcele nokubika ngokwethembeka. Lokhu kusho ukuthi ukuthuthukiswa kwamakhono nemingcele yokuhambisana kwakhiwe ngesikhathi esisodwa enqubweni efanayo yokuqeqeshwa kwangemva kokuqeqeshwa.
Ukuqeqeshwa kokuhambisana nobubanzi: Imodeli yaqeqeshwa ngokukhethekile ukuze icabange imingcele yomsebenzi, ihlukanise phakathi kwezinsizakalo ezigunyaziwe nezinsizakalo ezisizayo ezibonakala zingaxhashazeka kodwa ezingaphandle kobubanzi. Isiqeshana sokulandelana komcabango esinikezwe kuwebhusayithi (“don’t explore out of scope credentials”) sibonisa ukuthi ukwahlulela imingcele sekuyingxenye yokucabanga kwemodeli: iqala ngokunquma ububanzi, bese ikhetha isenzo.
Ukuphindaphinda okuqhutshwa ukuhlola: I-OpenAI yazakhela izivivinyo zangaphakathi eziningi ukuze ithuthukise amakhono. I-ExploitGym ihlola ukuxhashazwa kobuthakathaka obunqunyiwe, i-RemoteExploitBench ihlola ukuhlasela okukude kwezinhlelo zangempela, i-ExploitGym Honeypot ihlola ukuthambekela kokudlula imingcele, kanti i-SRE-Bench ihlola ukuhlaziya ngokuhlehlisa. I-RemoteExploitBench ne-Honeypot kwakuyizivivinyo ezintsha ezakhelwe lolu hlobo lwemodeli, futhi ukuthuthukiswa kwemodeli nokwakhiwa kwezivivinyo kwahamba kanyekanye.
Uhlelo lokuvikela ngesikhathi sokusetshenziswa: Kanye nokuvula amakhono, kusetshenziswa imingcele eminingi yesikhathi sokusebenza, okuhlanganisa ukuqapha kokungazinzi kwenethiwekhi, ukuqapha okuqondiswe ezenzakalweni ezithile zobuthakathaka, i-Private Safety Processing, uhlu olumhlophe lwama-IP oluzenzakalelayo lwe-Daybreak Red nokuqapha kokuvuza kweziqinisekiso, kanye nokuvimbela okuqinile kwezicelo ezinobungozi obukhulu ezimbonini zokulawula, okhethweni nakwezinye izindawo.

Qinisekisa ukufaneleka: Qinisekisa ukuthi ungumvikeli noma umcwaningi wezokuphepha ogunyaziwe, ngoba i-GPT-6 Cyber itholakala kuphela kumakhasimende e-Daybreak Red agunyaziwe ngemva kokubuyekezwa.
Thumela isicelo: Kuwebhusayithi ye-OpenAI ku-https://openai.com/zh-Hans-CN/daybreak/ faka isicelo sohlelo lwe-Daybreak, khetha izinga le-Red bese uchaza isimo sokusetshenziswa nenjongo ehleliwe.
Qedela ukubuyekezwa: Izinhlangano kufanele zilungiselele ukungena kanye, ukuqinisekiswa kwezinto eziningi kanye nohlu olumhlophe lwama-IP; abantu ngabanye kufanele bamenywe, baqinisekiswe ngobunikazi futhi basebenzise ukhiye wokuphepha wehadiwe.
Thola ukufinyelela: Ngemva kokugunyazwa, sebenzisa imodeli nge-API noma nge-Codex; ungafaka nesicelo sesiteshi esikhethekile sengqalasizinda ebalulekile.
Setha ububanzi: Chaza ngokucacile imingcele yokugunyazwa nemithetho yokuzibandakanya emsebenzini ngamunye; imodeli izoqala ngokucabanga ngobubanzi ngaphambi kokuthatha isinyathelo.
Yenza ngaphakathi kwe-sandbox: Sebenzisa indawo ehlukanisiwe ekutholeni ubuthakathaka, ekuthuthukiseni ama-exploit noma ekuhloleni ukungena; imisebenzi enobungozi obukhulu idinga ukuqondiswa ngumuntu.
Buyekeza imiphumela: Hlola lokho okutholwe yimodeli, ubufakazi neziphakamiso zama-patch, bese ukusebenzisa ekulungiseni noma ekubikeni kuphela ngemva kokuqinisekisa ukunemba.

Umehluko omkhulu emandleni: I-ExploitGym ithole u-48.5%, okuwukukhula okuhlobene ngo-51% uma kuqhathaniswa ne-GPT-5.6 Cyber, okubonisa ukweqa okucacile kwesizukulwane samakhono.
Ukuphumelela ekuhambisaneni nemingcele: Ngaphandle kwesaziso, ayidluli imingcele ngo-99.93%; uma inikezwe isaziso, igcina ububanzi bokugunyazwa ngo-100%, ixazulula ukungqubuzana phakathi kokuvula amandla okuhlasela nokuwalawula.
Imiphumela yangempela eqinile: Ama-exploit angu-948 okuthola amalungelo ku-Linux kanye nobuthakathaka obungaziwa obuphezulu obungu-81 ku-Firefox kuyimiphumela yangempela ephindiwe futhi edaluliwe.
Amamodeli ezizukulwane eziningi asenokwanda kwamakhono: Ku-Firefox, oseskenwe kaningi ngamamodeli aphambili ezizukulwane eziningi, isathole ama-0day angu-81.
Umjikelezo ophelele kusukela ekutholeni kuya ekulungiseni: Ingathola ubuthakathaka, ibhale i-exploit futhi ikhiqize i-patch ephakanyisiwe, ihlanganise lonke uchungechunge “lokuthola—ukuqinisekisa—ukulungisa”.
Inzuzo kwezomnotho zokuhlasela nokuvikela: Izindleko ze-API ze-exploit ngayinye esebenzisekayo cishe zingamadola angu-1,279, ngaphansi kakhulu kwemiklomelo yemiphumela efanayo emincintiswaneni efana ne-Pwn2Own, engamadola angu-20,000–70,000.
Ingozi yokudlula imingcele iyalawuleka: Uma iqhathaniswa nezinga lokudlula imingcele le-honeypot eliseduze nengxenye ku-GPT-5.6 Sol, i-Cyber yehlise lokhu kuziphatha ezingeni elingalawulwa ngobunjiniyela.

Iwebhusayithi yephrojekthi: https://openai.com/index/introducing-gpt-6-cyber/

Ukuhlolwa kokuphepha kwamakhodi ezinkampani: Amaqembu ezokuphepha ayisebenzisa ukuhlola amakhodi amakhulu, ukuthola ubuthakathaka, ukuqinisekisa umthelela nokukhiqiza izixazululo. I-Cisco, i-Palo Alto, i-SentinelOne nezinye sezike zayisebenzisa ocwaningweni lobuthakathaka nasekuhlolweni kwamakhodi.
Ukugcinwa kwesofthiwe yomthombo ovulekile: Kumaphrojekthi avuthiwe afana ne-Linux kernel ne-Firefox, isetshenziswa ekutholeni ubuthakathaka obungaziwa ngezinga elikhulu nasekukhiqizeni ama-patch aphakanyisiwe; sekukhona ubuthakathaka obubalulekile be-Linux obungu-304 obuhambisane nama-patch futhi bathunyelwa kubagcini bamaphrojekthi.
Amaqembu abomvu agunyaziwe nokuhlolwa kokungena: Amakhasimende e-Daybreak Red angayisebenzisa ukwenza uchungechunge oluphelele lokuhlasela ezindaweni zangempela ezifana ne-GitLab, PostgreSQL ne-Jenkins, ukuze ahlole ukusebenza kohlelo lokuvikela.
Ukuvikela ingqalasizinda ebalulekile: Abavikeli be-OT/ICS ezikhungweni zikagesi, ezitshalweni zamanzi nakwezinye izindawo bangafaka isicelo sesiteshi esikhethekile. Ekuhlolweni kwe-Dragos, yake yathola ngokuzimela ama-0day amabili kusofthiwe yobunjiniyela futhi yafunda indlela yokukhubeka kwe-breaker.