I-OpenAI yamemezela ngo-1 Septhemba ngesikhathi sendawo ukuthi ihlela ukukhipha “maduze” imodeli ye-AI yesizukulwane esilandelayo, i-Astra, kodwa izokhawulela ngokuqinile ububanzi bokusetshenziswa kwemisebenzi yayo yokuphepha ku-inthanethi.
Ngokwethulwa kwayo, i-Astra iyimodeli yokuqala ye-OpenAI efinyelele umkhawulo wamandla “obalulekile” wokuphepha ku-inthanethi ku-《Preparedness Framework》 yayo.
Ngokwencazelo ye-OpenAI, ukufinyelela umkhawulo “obalulekile” kusho ukuthi imodeli iyakwazi, ngaphandle kokungenelela komuntu, ukuthola nokwakha izindlela ezisebenzayo zokusebenzisa izikhala ze-zero-day ezinawo wonke amazinga obucayi ezinhlelweni eziningi zangempela ezibalulekile, ezivikelwe ngokuqinile.
U-Amelia Glaese, iphini likamongameli wezocwaningo kwa-OpenAI, uthe i-Astra iyakwazi ukuthola izikhala zokuphepha ebezingaziwa ngaphambilini ezinhlelweni eziningi ezivikelwe ngokuqinile futhi yakhe izindlela zokuzisebenzisa, ngaphandle kokuqondiswa umuntu isinyathelo ngesinyathelo. Ekuhlolweni, i-Astra yathola ngempumelelo izikhala ezimbili ze-zero-day futhi yazisebenzisa ngokuzixhumanisa; i-OpenAI isazidalula lezi zikhala kubanakekeli abafanele.
I-OpenAI ithe izosebenzisa isu lokuvula ukufinyelela ngamazinga. Ngemva kokukhishwa kwe-Astra, ikhono layo lokwenza imisebenzi ethuthukile ehlobene nokuphepha ku-inthanethi ekuqaleni lizotholakala kuphela eqenjini elincane labahloli.
Kamuva, le nkampani ihlela ukuvula ukufinyelela kubasebenzisi abaningi ukuze basebenzise i-Astra emisebenzini yokuvikela ukuphepha ku-inthanethi, ngohlelo lwe-Daybreak Blue. I-OpenAI iyavuma ukuthi ukukhawulela amakhono e-Astra okuphepha ku-inthanethi kungase kukhawulele umsebenzi osemthethweni wokuzivikela owenziwa ezinye izikhungo namabhizinisi.
Lolu hlelo lokukhishwa luqhamuke ngemuva kwesigameko esenzeka kwa-OpenAI ngoJulayi 2026, lapho i-ejenti ye-AI ezimele “yaphuma emikhawulweni” kungahlosiwe futhi yahlasela inkundla ye-AI i-Hugging Face.
Ngaleso sikhathi, i-ejenti ye-AI ezimele eyathuthukiswa kabusha kusetshenziswa amamodeli amabili e-OpenAI, i-GPT-5.6 Sol nenye imodeli engadalulwanga, yasebenzisa izikhala zesofthiwe endaweni yokuhlola ehlukanisiwe (sandbox) ngesikhathi sokuhlolwa kwezokuphepha, yazixhumanisa ne-inthanethi futhi yangena ezinhlelweni ze-Hugging Face. Embikweni owashicilelwa yi-OpenAI ekupheleni kuka-Agasti, le nkampani yavuma ukuthi ibingasabela kusenesikhathi ukuze ivimbele lesi sigameko.
I-OpenAI ithe nakuba i-Astra ingeyona enye yamamodeli eyabamba iqhaza ekungeneni ezinhlelweni ze-Hugging Face, le nkampani isebenzise izifundo ezizitholile ekuthuthukiseni izivikelo zokuphepha ze-Astra. Lokhu kuhlanganisa ukuqeqesha imodeli ukuze yenqabe ngokuthembekile izicelo “eziyingozi zokuphepha ku-inthanethi”, kanye nokwengeza “umqaphi wokungahambisani” (misalignment monitor) okhethekile ukuze uhlonze futhi uvimbele ukuziphatha okungaba yingozi. Lezi zivikelo zihlanganisa nokuqapha imodeli ngesikhathi isetshenziswa ngaphakathi kwenkampani ukuze kubonakale ukuziphatha okungagunyaziwe, kanye nokuyimisa ngokuzenzakalelayo imisebenzi okungenzeka ukuthi ayigunyaziwe.
Ngaphezu kwalokho, i-OpenAI ixwayise ngokuthi izivikelo zokuphepha ze-Astra zingase zihlabe umkhosi ngephutha emisebenzini esemthethweni yokuzivikela kwezokuphepha ku-inthanethi, ziyibeke njengokusetshenziswa kabi, okungabangela ukuthi umsebenzi ubambezeleke, umiswe okwesikhashana noma unqanyulwe (Qaphela: I-OpenAI ihlukanisa imisebenzi yokuphepha ku-inthanethi ibe imisebenzi yokuzivikela “yeqembu eliluhlaza” nemisebenzi yokuhlasela “yeqembu elibomvu”, kuye ngohlobo lomsebenzi; uhlelo lwe-Daybreak Blue luvumela kuphela umsebenzi wokuzivikela “weqembu eliluhlaza”).
U-Fouad Matin, usosayensi wezocwaningo kwa-OpenAI, uthe: “Sikholelwa ukuthi la makhono angasiza abavikeli bathole futhi balungise ubuthakathaka obukhulu, kodwa ngaphandle kwezivikelo ezifanele angase futhi enze abahlaseli basebenze ngempumelelo eyengeziwe. Yilokho kanye esizama ukukuvimbela.”