Umnyango wezokuphepha ku-inthanethi we-Google, i-Wiz, umemezele uhlelo lwama-AI agent lokuthola ubungozi i-Atlas, olungakwazi ukuzitholela ubungozi ngokuzimela. Kuze kube manje, i-Atlas isithole ubungozi bokuphepha obungaphezu kuka-200, futhi yazuza nomklomelo ka-GitHub wokuthola ubungozi ongu-US$100,000 (qaphela: ngokwezinga lokushintshisana lamanje, cishe u-¥678,000).
Ngokwethulwa kwayo, i-Atlas yakhiwe ngama-AI agent amaningi, lapho i-agent ngayinye ibhekele umsebenzi ohlukile. Indlela esebenza ngayo iyonke ifana neqembu labacwaningi bezokuphepha. Ama-agent ahlukene abhekele ukuqonda i-codebase, ukuphakamisa imibono yobungozi, ukuqinisekisa nokuphikisa imibono, bese ekugcineni eqinisekisa futhi abike ubungozi.
Ngokungafani nemodeli eyodwa ye-AI evamile eyenza umsebenzi kusukela ekuqaleni kuya ekugcineni, i-Atlas ihlukanisa umsebenzi ube yizigaba eziningi futhi isebenzisa ukubambisana okuhleliwe ngohlelo ukuze ithole ubungozi. Okokuqala yakha imodeli yezinsongo ye-codebase yesoftware eqondiwe, bese isebenzisa i-Code Property Graph ukuhlaziya ukwakheka kohlelo nokuthola izindlela ezingase zisetshenziselwe ukuhlasela.
Ngemva kwalokho, ama-AI agent amaningi ohlelweni aphakamisa imibono yobungozi ngokuhlukana futhi ayiqinisekise. Akha iziphakamiso ezingase zibe khona zobungozi obuhlanganisa izinyathelo eziningi, bese lezi ziphakamiso zinikezwa ama-agent amathathu anendima ethi “ukusekela impikiswano”, “ukuhlaziya ukuphikisa” kanye “nokwenza isinqumo sokugcina”. Kugcinwa kuphela ubungozi obuphumelele ekuqinisekisweni.
Ekugcineni, i-Atlas izakhela ngokuzenzakalelayo indawo yokuhlola ezinikele, ifake izingxenye ezidingekayo, bese isebenzisa okokufaka kokuhlasela ukuze ivuse ubungozi futhi iqinisekise ukuthi bukhona ngempela.
Ukuze yehlise izindleko zokusebenza, i-Atlas isebenzisa isu elinemininingwane eminingi lokwabela i-AI imisebenzi. Isibonelo, ezinyathelweni ezidinga amandla aphezulu okucabanga, njengokuhlaziya uchungechunge oluyinkimbinkimbi lokuxhaphaza ubungozi kanye nokwenza isinqumo sokugcina ngobungozi, uhlelo lubiza imodeli enkulu enamandla kakhulu. Kodwa emisebenzini echazwe ngokucacile nephindaphindeka kakhulu, njengokuhlaziya i-Code Property Graph nokuhlunga ubungozi esigabeni sokuqala, kusetshenziswa amamodeli amancane anezindleko eziphansi.
Ngaphezu kwalokho, i-Atlas ihlanganisa ubuchwepheshe bokuhlaziya ikhodi obunemiphumela enqunyiwe ukuze iqondiswe kahle nangempumelelo. Ihlunga kusenesikhathi imiphumela engamanga ngohlelo lokubuyekeza oluphikisanayo, bese isebenzisa i-Harness, uhlaka lokuqalisa olwenzelwe ucwaningo lobungozi, ukuze iqinisekise imiphumela futhi iphinde yehlise izindleko zokusebenza.
I-Wiz ithe lo mklamo wenza i-Atlas ikwazi ukwenza ukuhlaziya okujulile kwekhodi ngezinga elikhulu nangezindleko eziphansi. Uhlelo luphinde luzikhiqizele ngokuzenzakalelayo uhlelo lokuqinisekisa ukuhlasela kwebungozi ngakunye olutholakele, ngaleyo ndlela lwehlise amathuba emibiko engamanga. Ngokwemiphumela yohlolo oluyisisekelo locwaningo lobungozi be-AI i-CyberGym, olucashunwe yi-Wiz, i-Atlas ithole u-90.9%, yadlula amamodeli ancintisanayo afana ne-GPT-5.5-Cyber, i-Claude 4.6 kanye ne-Claude Mythos Preview.
Mayelana nokusetshenziswa kwangempela, i-Wiz ithe ngemva kwezinyanga ezimbalwa zokuthuthukiswa, i-Atlas isivele ithole futhi yaqinisekisa ubungozi obungaziwa ngaphambilini obungaphezu kuka-200 kumaphrojekthi amaningi adumile avulekile. Iphinde yasiza i-GitHub ukuthola ubungozi obukhulu bokusebenzisa ikhodi ukude, i-CVE-2026-3854, futhi ngenxa yalokho yathola umklomelo ka-US$100,000 (ngokwezinga lamanje lokushintshisana, cishe u-¥678,000).