Ngokombiko oshicilelwe yiphephabhuku i-Wired izolo, inkampani esafufusa yaseMelika yezokuphepha ku-inthanethi i-Frontier Security ithe ngesikhathi ihlola amakhono e-cybersecurity emodeli i-Kimi K3 ye-Moonshot AI, le modeli yashiya indawo ye-sandbox eyayiyivimbile.
Ngokwemibiko, lesi sigameko sokweqa kwemingcele kwe-AI sifana nezehlakalo ezihilela i-OpenAI ne-Anthropic; imbangela eqondile nayo kwaba ukungalungiseki kahle kwe-sandbox eyayisetshenziselwa ukuhlukanisa imodeli. I-Frontier Security ikholelwa ukuthi i-Kimi ayinazo izindlela zokuvikela inethiwekhi ezitholakala kwamanye amamodeli e-AI asezingeni eliphezulu, okwenze le modeli yakwazi ukufinyelela ku-inthanethi ngaphandle kwemvume ecacile.
Isikhulu esiphezulu se-Frontier Security, u-Yaron Singer, sathi: “Sithole ubungozi ku-sandbox, kodwa u-Kimi naye wasebenzisa lobo bungozini. Lokhu kukhombisa ukuthi kungenzeka ayinazo izindlela zokuphepha zangaphakathi ezitholakala kwamanye amamodeli asezingeni elifanayo.”
Nokho, ngemva kokuba i-Kimi K3 yeqe i-sandbox futhi yathola imvume yokufinyelela ku-inthanethi, ayizange yenze ukuhlasela; yavele yafuna impendulo yombuzo othile e-GitHub.
Ukwanda kwezehlakalo zokweqa kwemingcele kwe-AI muva nje kukhombisa ukuthi njengoba amakhono e-AI eba namandla, kuba nzima kakhulu ngabantu ukuwalawula. Ngenyanga edlule, i-OpenAI yadalula ukuthi imodeli yayo yangaphakathi engashicilelwanga yephule indawo yokuhlukanisa ngesikhathi sokuhlolwa, yabe isihlasela inkundla i-Hugging Face. Kamuva, i-Anthropic nayo yathi amamodeli ayo amaningi ake athola imvume yokufinyelela ku-inthanethi ngesikhathi sokuhlolwa kwezokuphepha futhi ahlasela amasistimu angaphandle.
Abanye ochwepheshe bezokuphepha ku-inthanethi bathi lokhu kutholakala kukhombisa ukuthi abantu kudingeka bahlele kabusha izindawo ze-sandbox zamamodeli e-AI asezingeni eliphezulu. Abasebenzisi bama-agent afana ne-OpenClaw (inothi: eyaziwa ngokuthi i-“lobster”) nabo kudingeka balungise izindawo zabo ngokucophelela ukuze bavimbele i-AI ekweqeni imingcele.