IAtlasi yeNgxoxoGPTIsikhangeli se-OpenAI esisebenza nge-AI siye saba yinxalenye ephambili yengxoxo yokhuseleko lwedijithali njengoko sifumana iimpawu zewebhu ezizimeleyo ngakumbi, ezifana notshintsho lweqonga olwenzeka nge Ivenkile yeeapps zeChatGPTEsi sixhobo sithembisa ukwenza imisebenzi yemihla ngemihla ibe lula njengokufunda ii-imeyile, ukuzalisa iifomu, okanye ukuhamba phakathi kwamaphepha ahlukeneyo, kodwa loo mandla afanayo ayenze yaba yindawo enomtsalane yokuhlaselwa ngokukhawuleza ngenaliti.
Ngenxa yesi simo, inkampani ekhokelwa nguSam Altman ibhengeze ukomelezwa okukhulu kokhuselo lweChatGPT Atlas ukulwa neendlela ezifuna ukufaka imiyalelo enobungozi kumxholo obonakala ungenabungozi. I-OpenAI iyavuma ukuba isoyikiso asiyi kunyamalala, kodwa igxininisa ukuba ingenza njalo kwandisa kakhulu ubunzima kunye neendleko kwezi hlaselo, into ebalulekileyo kubasebenzisi ngabanye kunye nemibutho eSpain nakwamanye amazwe aseYurophu, ingakumbi kwiindawo ezixhomekeke kwizivumelwano zelifu ezifana neso sisayinwe ne Amazon.
Yintoni i-prompt injection kwaye kutheni ibangela umngeni kwi-agent mode?
Umnxeba inaliti okanye umyalelo wokufakwa ngokukhawuleza Iye yaba yenye yezona zinto zibalulekileyo ezibangela ubuthathaka kwiinkqubo ze-AI ezivelisayo. Indlela yokusebenza ilula kakhulu: umhlaseli Ifihla imiyalelo enobungozi kwii-imeyile, amaphepha ewebhu, amaxwebhu, okanye kwanamaqhekeza angabonakaliyo., sinethemba lokuba imodeli yolwimi iya kuzitolika njengemiyalelo ekufuneka ilandelwe.
Kwimeko I-ChatGPT Atlas kunye nemo yearhente yayoIngxaki iyanda kuba isikhangeli senzelwe hlalutya umxholo owenziwe ngabantu besithathu kwaye wenze izinto ngokuzimeleyoUngandwendwela iiwebhusayithi, ufunde imiyalezo, ugcwalise iifomu, okanye uqalise imisebenzi enzima ngaphandle kokuba umsebenzisi ahlole inyathelo ngalinye ngesandla, nto leyo evula ucango lwemiyalelo efihlakeleyo ekhokelela kuyo. izenzo ezingafunwayo.
I-OpenAI ichaze ukuba imo yearhente inako ukwenza sebenza ngamanyathelo amaninzi okanye amakhulu ukugqiba umsebenzi ocelwe ngumsebenzisi. Ukuba kufakwa i-prompt injection eyilwe kakuhle phakathi kwaloo nkqubo, i-AI ingaphela bephula imiqobo yabo yokhuseleko kunye nokuphumeza imiyalelo ebeya kuvalelwa ngokuqhelekileyo.
Phakathi kwezinto ezibangela ingxaki enkulu kule nkampani kukho ukufakwa kwebhodi yokunamathisela, indlela apho inkqubo ikopisha ngokuzenzekelayo ikhonkco okanye umxholo onobungozi ngaphandle kokuba umntu ophambi kwekhompyutha aziIngozi ivela xa umsebenzisi encamathisela loo mbhalo kwi-address bar okanye kwenye i-application, apho uhlaselo lusebenza khona.
I-OpenAI ngokwayo ibeka i-prompt injection kudidi olufanayo nolwalo ubuqhetseba kwi-intanethi okanye ubunjineli bezentlaloEzi zizinto ezinokuthi zincitshiswe, kodwa kunzima ukuzisusa ngokupheleleyo. Yiyo loo nto ndichaza ezi ntlobo zohlaselo njenge umngeni wesakhiwo wexesha elide kuyo nayiphi na iarhente ye-AI ehambahamba kwiwebhu evulekileyo.
Uhlaziyo lokhuseleko: ukhuselo oluqhubekayo kunye nokuphendula ngokukhawuleza
Ukujongana nale meko, i-OpenAI iqalise uhlaziyo oluthile lokhuseleko lwe-ChatGPT Atlasigxile ekubhaqweni kwangethuba kunye nokunciphisa ukuhlaselwa ngenaliti. Eyona nto iphambili kolu qiniswayo kuku imodeli entsha eqeqeshwe ngokukodwa ukujongana neentshaba loo mzamo wokulawula indlela iarhente eziphatha ngayo.
Le modeli idityaniswe kwi inkqubo yokuzikhusela eqhubekayoyenzelwe ukulungisa ukhuseleko lwesiphequluli njengoko kuvela iindlela zokuhlasela ezintsonkothileyo. Le nkampani ithi injongo kukuba ukufumanisa nokulungisa ubuthathaka bangaphakathi ngaphambi kokuba zibe “zizixhobo ezisebenzayo”—oko kukuthi, ngaphambi kokuba abahlaseli bazisebenzise kwiindawo zokwenyani. Olu luhlu lomsebenzi luhambelana neziseko zophuhliso kunye namanyathelo okhuseleko aqhutywa ngamaqabane afana Umanyano lweSamsung kunye ne-OpenAI.
Enye into ebalulekileyo kukuphunyezwa umjikelo wokuphendula ngokukhawulezaYaphuhliswa ngentsebenziswano neqela langaphakathi leRed Team le-OpenAI. Eli qela lizinikele kwi phanda iivektha ezintsha zohlaselo, uzivavanye kwiindawo ezilawulwayo, kwaye usebenzise iindlela zokunciphisa umngcipheko ngobuchule obukhulu obunokwenzeka, ngendlela amaqela okhuseleko lwe-intanethi asebenza ngayo kwiinkampani ezininzi ezinkulu zobuchwepheshe.
Enyanisweni, oku kuguqulela kwi I-ChatGPT Atlas ifumana uhlaziyo oluhlala lusenziwa olujolise ekuphenduleni ngononophelo ngakumbi xa kukho iindlela ezingaqondakaliyo: ukusuka kwimiyalelo ephikisanayo ekwisiqendu ukuya kwimiqondiso engabonakaliyo esasazeke kwiphepha lewebhu okanye kuthotho lwee-imeyile.
I-OpenAI igxininisa ukuba eli cebo alisosisombululo sexeshana, kodwa inkqubo eqhubekayo eya kuhamba nesikhangeli njengoko inqanaba laso lokuzimela lisandaLe mbono ibaluleke kakhulu kwiinkampani zaseYurophu, eziqwalasela kakhulu uzinzo, ukuthotyelwa kwemithetho, kunye nolawulo lomngcipheko xa zifaka izisombululo ze-AI kwimisebenzi yazo yokusebenza.
"Umhlaseli ozenzekelayo" ofunda njengo-hacker
Enye yezona zinto zibalaseleyo kwindlela ye-OpenAI kukudala "umhlaseli ozenzekelayo osekelwe kwi-LLM"I-bot eyilelwe ukudlala, ngendlela elawulwayo, indima yomqhekezi okhangela ubuthathaka kwinkqubo. Kunokuba ithintelwe kuvavanyo oluzinzileyo, lo mhlaseli wokwenziwa funda kwaye uhlengahlengise amaqhinga akho ixesha elidlulele.
Le nkampani ichaza ukuba i-bot iqeqeshwe yi ukomeleza ukufundaLe yindlela apho inkqubo ifumana impendulo ngokusekelwe ekubeni imizamo yayo yokuhlasela iphumelele okanye ayiphumeleli. Xa iarhente yeChatGPT Atlas ixhathisa uhlaselo, umhlaseli uhlalutya impendulo, alungise icebo lakhe, aze Zama kwakhona ngokulandelelana.
Ngokwedatha eyabelwe yi-OpenAI, lo mhlaseli ozenzekelayo unako ukwenza yenza iarhente ukuba yenze imisebenzi eyinkimbinkimbi kakhulu neyingozienokudlula amashumi okanye amakhulu amanyathelo adibeneyo. Injongo ayikokuba olu hlaselo lufikelele kumsebenzisi wokugqibela, kodwa ukuphinda uvelise kwiimeko zelebhu ezinokwenzeka kwihlabathi lokwenyani.
Zonke ezi zilingo zenzeka iindawo ezilinganisiweyoukuze inkampani ikwazi ukubona ngokweenkcukacha indlela iarhente ephendula ngayo kwimizamo nganye yokukhohlisa. Eli nqanaba lokubonakala livumela chonga iipatheni zokuziphatha ezinengxaki kwaye uqinise ukhuselo kwiindawo ezithile ekunokuba nzima ukuzibona ngokusebenzisa uvavanyo lwezandla okanye uhlaselo lwangaphandle kuphela.
I-OpenAI ithi ngenxa yale nkqubo ifezekisa fumana amaqhinga okuhlasela angazange abonwe ngaphambiliOko kukuthi, iindlela ezingakhange zivele kwizenzo zokusebenzisa amaqela abomvu okanye kwiingxelo zomntu wesithathu. Olu buchule bokuhlala inyathelo elinye phambi kwabahlaseli abanokubakho, ngokutsho kwenkampani, yenye yezona nzuzo ziphambili zokudibanisa iimodeli zolwimi neendlela zokhuseleko eziphambili.
Imizekelo yokwenyani: ukusuka kwii-imeyile eziguqulweyo ukuya kwiikhonkco ezikhutshelweyo ngokungazi
Ukubonisa impembelelo ebonakalayo yolu phuculo, i-OpenAI ibonise imizekelo ye Indlela iChatGPT Atlas eyayiziphatha ngayo ngaphambi nasemva kohlaziyoKwelinye lamatyala akhankanyiweyo kakhulu, umhlaseli ufaka umyalelo ofihliweyo kwi-imeyile eyalela iarhente thumela umyalezo kwi-CEO yenkampani engeyonyani echaza ngokushiya emsebenzini komsebenzi owayelixhoba lohlaselo.
Kwiinguqulelo zangaphambili zenkqubo, imo yearhente Walandela umyalelo ngaphandle kokubuza imibuzo emininzi.kuba itolike umxholo njengomsebenzi osemthethweni ovela kumsebenzisi. Emva kokwaziswa kwezikhuselo ezintsha, isikhangeli ifumanisa ukuba ngumyalelo onobungozi ofihliweyo kwaye ukhetha ukulumkisa umsebenzisi endaweni yokuthumela i-imeyile.
Olu hlobo lwemiboniso lusebenza ukubonisa indlela ibhloko elula yombhalo efakwe kumyalezo wesiqhelo Oku kunokubangela iziphumo ezichaphazela kakhulu ukuba inkqubo ayinazo iindlela ezithile zokucoca nokubuza imibuzo ngee-odolo ezifunyenweyo.
Kwangaxeshanye, inkampani ikhumbule ezinye iziganeko, ezifana nezo zinxulumene ukufakwa kwebhodi yokunamathiselaapho i-AI igqibele ikopisha amakhonkco akrokrisayo ngaphandle kolwazi lomsebenzisi. Ngolu hlobo lutsha lokhuseleko, injongo kukuba I-Atlas ichonga kwaye ivale ukuziphatha okungaqhelekanga kuloo ngcelele yezenzongaloo ndlela kunciphisa umda wokuba uhlaselo luqhubeke.
Kwimeko yaseYurophu, apho imithetho yokhuseleko lwedatha kunye nokhuseleko lwe-intanethi ingqongqo ngakumbi, la matyala okusetyenziswa asebenza njenge uhlobo lwendawo yokuvavanya ukuvavanya indlela izikhangeli ze-AI ezisebenza ngayo zinokudityaniswa kwiindawo zoshishino ngaphandle kokunyusa izinga lomngcipheko ocingelwayo.
Umngcipheko ongapheliyo, kwaye onke amehlo ajonge eYurophu.
Kwiingxelo zayo, i-OpenAI iyayamkela ithoni elumkileyo neyokwenyaniLe nkampani iyavuma ukuba "akunakwenzeka" ukuba ukuhlaselwa ngokukhawuleza ngenaliti kunokupheliswa ngokupheleleyo, njengokuba kungengazo zonke iintlobo zobuqhetseba kwi-intanethi ezinokupheliswa. Ngokombono wabo, eyona nto iphambili ikukuba ukunciphisa umphezulu wohlaselo kunye nempembelelo enokubakho, endaweni yokulwela ukhuseleko olupheleleyo.
Olu xilongo luhambelana nezilumkiso ezivela Iiarhente zokhuseleko lwe-cyber zaseYurophuabaye babonisa kudala ukuba iinkqubo ze-AI ezivelisayo zineengozi ezimele zilawulwe rhoqo. Le ndlela ibandakanya ulawulo lobugcisa, imigaqo-nkqubo yangaphakathi ecacileyo, kunye noqeqesho lwabasebenzisiendaweni yokuthembela kuphela kumqobo wetekhnoloji oqinisekileyo.
Okwangoku, ezinye iinkampani ezinkulu kweli candelo, ezifana neGoogle okanye iAnthropic, ziqalile cinga ngokutsha ngoyilo lweearhente zayo ukubandakanya izikhuselo ukususela kwinqanaba loyilo. Ingcamango eqhelekileyo kushishino kukuba Ukuzimela kwezi nkqubo kufuneka kuhlale kukhatshwa ziibhuleki kunye nee-counterweights. ezithintela umonakalo xa kukho into engahambi kakuhle.
Iingcali zokhuseleko zithi umngcipheko kwiibhrawuza ezisebenzisa i-AI unokuqondwa njengo isimbuku somgangatho wokuzimela kwearhente kunye nokufikelela kwayo kwizixhobo ezibuthathaka (ii-imeyile, iiakhawunti ze-intanethi, izixhobo zokuvelisa, kwaneentlawulo). Kulo balo, i-ChatGPT Atlas kunye nezisombululo ezifanayo zikwindawo ebucayi kakhulu kwiinkampani zaseYurophu ezisingatha idatha ebalulekileyo.
Le nyaniso inyanzela ababoneleli kunye nabasebenzisi ukuba gcina isimo sengqondo sokuthandabuza okunempilo: sebenzisa ithuba lokusebenzisa oomatshini, ewe, kodwa kuphephe ukudlulisela ngokungaziyo izigqibo ezinokuba nemiphumo yezomthetho, yezemali okanye yodumo kwiManyano yaseYurophu.
Amacebiso okusebenzisa ngokukhuselekileyo abasebenzisi kunye nemibutho
Kunye nophuculo lobuchwephesha, i-OpenAI yabelane ngayo Uthotho lweengcebiso zokusebenzisa i-ChatGPT Atlas ngokukhuselekileyo ngakumbiyenzelwe abasebenzisi ngabanye kunye neenkampani ezivavanya imo yearhente eSpain okanye kwamanye amazwe aseYurophu.
Okokuqala, inkampani iyacebisa ukunciphisa ukufikelela kwearhente kulwazi olubucayi ngakumbiOku kuthetha ukuthintela isikhangeli ekubeni sibe neemvume ezibanzi kwiiakhawunti ze-imeyile zeenkampani, iinkqubo zokuhlawula, okanye amaqonga angaphakathi ngaphandle kokuba kuyimfuneko. Ngale ndlela, nokuba ukufakelwa ngokukhawuleza kwenzeke ngempumelelo, impembelelo enokubakho iyancipha.
Ikwacebisa ukuba kuthathelwe ingqalelo izicelo zokuqinisekisa ngokucacileyo eziboniswa yinkqubo ngaphambi kokuba zenze amanyathelo afanelekileyo. Ukuphonononga ngononophelo ezi zilumkiso nokungazizamkeli ngokuzenzekelayo kuvumela umsebenzisi ukuba alawule. umgca wokugqibela wokuzikhusela xa kukho iindlela zokuziphatha ezikrokrisayo ezinokuthi imodeli ngokwayo ingazicocanga ngokupheleleyo.
Esinye isikhokelo kukunika iarhente imiyalelo ecacileyo nemfutshaneEndaweni yemisebenzi eqhelekileyo efana nokuthi “lawula yonke i-imeyile yam” okanye “lawula imali yam kwi-intanethi,” ngokunciphisa ububanzi bomsebenzi, iba nempumelelo ngakumbi. kunzima ngakumbi ukuba umxholo onobungozi utshintshe ngokupheleleyo injongo yokuqala yomsebenzi onikiweyo.
Okokugqibela, i-OpenAI icebisa ukusebenzisa imo yearhente ngcono kwiindawo apho umsebenzisi engangenanga khona Okanye ubuncinane yahlula ngokucacileyo iimeko ezibuthathaka kwezo apho kusetyenziswa khona iimpawu zesikhangeli eziphambili. Olu hluhlu, oluqhelekileyo kwiindlela ezilungileyo zokhuseleko, lunceda ukuthintela ukuba ubuthathaka obunokubakho bungasasazeki kuzo zonke iiakhawunti kunye neenkonzo.
Amanyathelo abhengezwe yi-OpenAI abonisa ukuba Uphuhliso lwe-ChatGPT Atlas lubandakanya ukufumana izakhono kunye nokukhusela indlela esebenza ngayo. Xa kujongwa imizamo yokulawula, ukuhlaselwa ngokukhawuleza ngenaliti kuya kuhlala kukho, kodwa ukusetyenziswa kwezixhobo zokuzikhusela eziqhubekayo, ukusetyenziswa kwabahlaseli abazenzekelayo, kunye nokwamkelwa kweendlela ezilungileyo ngabasebenzisi kunokwenza isikhangeli sibe sisixhobo esivuthiweyo nesinokuthenjwa, esilungiselelwe ukusetyenziswa ngokunzulu eSpain nakwamanye amazwe aseYurophu, ngaphandle kokulahlekelwa yimbono yokuba ukhuseleko lobukrelekrele bokwenziwa lucelomngeni oluya kufuna uhlengahlengiso oluqhubekayo kwiminyaka ezayo.
