- Ukuphepha kwe-AI kuhlanganisa kokubili ukuvikela amamodeli kanye nedatha ekuhlaselweni kanye nokusebenzisa i-AI ukuqinisa ukuphepha kwe-inthanethi nokunciphisa izehlakalo.
- Izingozi zihlanganisa ukwehluleka kobuchwepheshe (ukufaka idatha engozini, izibonelo zokuphikisana, amaTrojan) nezici zesakhiwo ezifana nohlanga oluncintisanayo kanye nokuntuleka kokubusa.
- Izinhlaka ezifana ne-NIST AI RMF, i-OWASP Top 10 ye-LLM, kanye ne-MITRE ATLAS zisiza ekwakheni izilawuli eziphikisanayo kanye nokuhlolwa ezinhlelweni ze-AI ekukhiqizweni.
- Ukufaka izimpahla ze-AI ohlwini, ukunciphisa ukufinyelela, ukuqapha ukusetshenziswa, kanye nokwamukela i-AISPM kuyizinyathelo ezibalulekile zokusebenzisa i-AI ngenkathi kunciphisa indawo yayo yokuhlasela.
Ubuhlakani bokwenziwa bungene kukho konke : kusukela endleleni esisebenza futhi sixhumana ngayo kuya endleleni izinqumo zezezimali, zezokwelapha, nezezempi ezenziwa ngayo. Lokhu kuba khona okukhulu kunohlangothi oluhle kakhulu, kodwa futhi kunohlangothi olumnyama: uma lezi zinhlelo zehluleka, ziphathwa kabi, noma zisetshenziswa kabi, umonakalo ungaba mkhulu kakhulu, kokubili ezinkampanini kanye nasemphakathini wonke.
Eminyakeni yamuva nje, ukukhathazeka komphakathi ngezingozi ze-AI kuye kwanda kanye nokwamukelwa kwayo. Ucwaningo lwakamuva lubonisa ukuthi iningi labantu lizizwa lingakhululekile kunentshiseko ngokwanda kwayo, nokuthi izakhamuzi eziningi zikubona kunengqondo ukuthi i-AI ingabangela izehlakalo ezimbi uma ingalawuleki. Ngesikhathi esifanayo, ucwaningo lobuchwepheshe kuphela luhlala lugxile kakhulu ekwenzeni i-AI ibe namandla kakhulu, kanti ingxenye encane kuphela ezinikele ekwenzeni iphephe futhi iqine kakhudlwana.
Sisho ukuthini ngokuphepha kwe-AI futhi kungani kubaluleke kangaka?
Uma sikhuluma ngokuphepha kobuhlakani bokwenziwa, sibhekisela kusethi yezindlela, izimiso zobuchwepheshe, izilawuli zokuphepha kwe-cyber , kanye nezinhlaka zokubusa ezenzelwe ukuqinisekisa ukuthi izinhlelo ze-AI ziqinile, ziyabikezeleka, ziyakwazi ukumelana nokuhlaselwa, futhi zihambisana nezinhloso zabantu . Lokhu kufaka phakathi kokubili ukuvikelwa kwendabuko (ukuvimbela ukuhlaselwa kwe-cyber, ukuvuza kwedatha, noma ukuphathwa kabi kwemodeli) kanye nezinkinga ezibanzi ezifana nokubandlulula, ukuntuleka kokubonakala, kanye nemiphumela engaba khona yesistimu.
Kumabhizinisi kanye nezinhlangano zikahulumeni, ukuphepha kwe-AI kubalulekile ngoba izinqumo ezibaluleke kakhulu zidluliselwa kumamodeli : ukuhlolwa kwengozi yezezimali, ukubekwa phambili kwezokwelapha, ukukleliswa kwe-resume, ukuhlaziywa kwedatha enkulu, ukuzenzekela kwenqubo, kanye nokutholwa kwezinsongo. Ukwehluleka okukhulu, ukusetshenziswa kwangasese, noma okunonya kungaholela ekulahlekelweni kwezigidi, izinhlawulo zomthetho, ukulimala kwedumela, futhi, esimweni esibi kakhulu, izinsongo ekuphepheni kwezwe.
Ezingeni lezenhlalo, i-AI ingakhulisa ukungalingani kanye nokuxhashazwa uma ingalawulwa kahle. Izinhlelo ezingabonakali noma lezo eziqeqeshwe ngedatha ebandlululayo zingabandlulula amaqembu athile ngendlela ehlelekile, ziqinise ukubhekwa kwabantu abaningi, noma zilethe imikhankaso emikhulu yolwazi olungalungile. Futhi, uma sibheka phambili, amathuba okufinyelela ezinhlotsheni zobuhlakani obujwayelekile bokwenziwa (i-AGI) noma ngisho nobuhlakani obuphezulu (i-ASI) kuvule impikiswano mayelana nezingozi zokuphila uma lezi zinhlelo zingahambisani nezindinganiso zabantu futhi zibalekela ukulawula okusebenzayo.
Ochwepheshe abavumelani ngokuthi yiziphi izingozi eziphuthumayo kakhulu noma ezingaba khona , kodwa kukhona ukuvumelana okuthile kokuthi amathuba emiphumela emibi kakhulu awancane. Ucwaningo lwabacwaningi be-AI lulinganisele cishe amathuba angu-5% ezimo ezimbi kakhulu (isibonelo, ukuwa kwempucuko) okuvela ku-AI ethuthukisiwe. Abanye babheka lokhu kukhathazeka njengehaba, kanti abanye bakholelwa ukuthi ngenxa yokuthi kungathi sína kangaka, kuwukuhlakanipha ukuzilindela kusengaphambili futhi ungathembeli "ekulungiseni njengoba siqhubeka."
Kusukela ezixwayisweni zakuqala kuya kuhlelo lwesimanje lokuphepha lwe-AI
Nakuba kungase kubonakale njengesihloko samuva nje, izingozi zokwenza izinto ngobuhlakani ziye zaxoxwa ngazo kusukela ekuqaleni kokubala. Kusukela maphakathi nekhulu lama-20, abantu abanjengoNorbert Wiener baxwayisa ngokuthi, ngezinga esilinike imishini ekwazi ukufunda ngokuhlangenwe nakho, sacabanga nokuthi izoziphatha ngendlela ephambene nezifiso zethu.
Ngokuhamba kwesikhathi, lokhu kukhathazeka kwaba yinto ecacile. Ngasekupheleni kweminyaka yama-2000, izinhlangano zobungcweti ezifana ne-AAAI zakhuthaza izifundo ngomthelela wesikhathi eside wezenhlalo we-AI. Umyalezo ojwayelekile wawungabaza izimpikiswano eziyingozi kakhulu zezindaba zesayensi, kodwa ngesikhathi esifanayo wavuma isidingo sezindlela ezingcono zokuqinisekisa izinhlelo eziyinkimbinkimbi nokunciphisa imiphumela engalindelekile.
Ngawo-2010, impikiswano yaqina. Izincwadi ezinethonya mayelana nobuhlakani obuphezulu kanye nezingozi zokuphila zakhuthaza ukudalwa kwezikhungo ezintsha zocwaningo ezinikezelwe ku-AI "ezuzisayo nezuzisa abantu", futhi kwavela izinhlelo zokusiza umphakathi ezaxhasa amaphrojekthi athile ngokuphepha kobuchwepheshe, umthelela emphakathini, kanye nokuhambisana kwemigomo. Ngesikhathi esifanayo, abantu abadumile njengo-Elon Musk, uBill Gates, noStephen Hawking baqala ukukhuluma ngokukhululekile ngezingozi ezinkulu uma i-AI ithuthuka ngaphandle kwezivikelo.
Kwahlelwa izingqungquthela zomhlaba wonke, kwashicilelwa ama-ajenda ocwaningo "ngezinkinga ezithile zokuphepha kwe-AI", kwathi kwasungulwa izimiso ezifana nalezo ze-Asilomar, ezazihlanganisa izincomo ezicacile: ukubambisana phakathi kwamaqembu ukugwema ukwehlisa amazinga okuphepha, ukuhlaziywa komthelela, ukucaca lapho kungenzeka khona, njll. Lo msebenzi wahlanganisa insimu yawo: ubunjiniyela bokuphepha be-AI , ngenhloso yokulindela ukwehluleka nokuklama izinhlelo eziphephe kakhulu ngangokunokwenzeka ngaphambi kokusetshenziswa kwazo ngobuningi.
Eminyakeni yamuva nje, ohulumeni nabalawuli bajoyine lo mzamo . Ama-workshop ahlelwe yi-White House, izinhlelo ezivela ezinhlanganweni zase-US (i-DARPA, i-IARPA, i-NSF), i-National AI Strategy yase-UK, kanye neziqondiso zokuziphatha zaseShayina zibonisa ukuthi ukuphepha kwe-AI akuseyona nje indaba yezemfundo, kodwa kuyindaba yenqubomgomo yomphakathi. EYurophu, uMthetho we-AI wethula imikhawulo ecacile ekusetshenzisweni okuthile (isibonelo, ukuqashelwa kobuso obukhulu noma ukuphathwa okungabonakali) futhi usungula izibopho ezengeziwe zezicelo "ezinobungozi obukhulu".
Izindawo zobuchwepheshe ezibalulekile: ukuqina, ukuqapha, kanye nokuqondanisa
Ucwaningo lokuphepha lwe-AI luvame ukuhlukaniswa ngezigaba ezintathu eziyinhloko: ukuqina, ukuqapha, kanye nokuqondanisa . Ngayinye ibhekana nohlobo oluhlukile lwengozi, kodwa empeleni zixhumene kakhulu.
Ukuqina kugxila ezinhlelweni ze-AI ezisebenza ngokufanele ezimweni eziningi ezahlukene, kufaka phakathi lezo ezingamelwe kahle kudatha yokuqeqeshwa. Lokhu kuhilela ukubhekana "nama-black swans" (izehlakalo ezingavamile kodwa ezilimaza kakhulu) kanye nokuhlasela lapho othile eguqula khona ngamabomu okokufaka ukuze aphoqe amaphutha.
Isibonelo esivamile sokuntuleka kokuqina lapho kubhekene nezenzakalo ezingavamile yi- 2010 Flash Crash , lapho izinhlelo zokuhweba ezenzakalelayo zasabela ngokungafanele ezimeni ezingavamile zemakethe futhi zaqeda cishe amadola ayizigidigidi ngemizuzu. Ezweni elingokoqobo, izimoto ezizimele ziyaqhubeka nokulwa "namacala asemakhoneni": izimpawu zomgwaqo ezingavamile, izimo zezulu ezimbi kakhulu, noma inhlanganisela yezinto ezingavamile ukuvela kudatha yangempela.
Ngesikhathi esifanayo, amamodeli amaningi okufunda okujulile abonakale esengozini enkulu yezibonelo eziphikisanayo : ukuphazamiseka okuncane ezithombeni, kumsindo, noma embhalweni, okungabonakali emehlweni omuntu, okubangela amaphutha ngokweqiniso ekuhlukaniseni. Lokhu kuthinta izinhlelo zokubona, ukuqashelwa kwenkulumo, kanye namamodeli olimi. Okwenza izinto zibe zimbi kakhulu, akuyona nje indaba yokukhohlisa imodeli "eyinhloko": wona kanye amamodeli omvuzo noma okuhlola asetshenziswa ukuqeqesha abanye angasetshenziswa ukuze akhulise amaphuzu ngokwenziwa ngaphandle kokuthuthukisa ukusebenza emsebenzini ohlosiwe.
Ukuqapha kubheka indlela yokuthola kusenesikhathi lapho uhlelo lwe-AI luhluleka, luhlaselwa, noma lusetshenziswa kabi. Kuhlanganisa ukulinganiswa kokuqiniseka (ukuqinisekisa amathuba okukhishwa kwemodeli okubonisa ngokunembile ukuthembeka kwayo), ukutholwa kokungasatshalaliswa (lapho uhlelo luhlangana nokuthile "okuhlukile kakhulu" kulokho olukubonile ngaphambili), kanye nokuqapha ukusetshenziswa okungenzeka kube yingozi.
Okokugcina, ukuhambisana kuqinisekisa ukuthi izinhlelo ezithuthuke kakhulu, ikakhulukazi lezo ezenza izinqumo ezithinta kakhulu noma ezikwazi ukwenza izinto ngokuzimela, ziphishekela izinhloso esizibona zamukelekile futhi azithuthukisi amasu ayingozi okuzifeza. Lokhu kuhilela imiqondo efana nokuhlangana kwezinsimbi (ukuthambekela kwama-ejenti athuthukile ekufuneni imigomo efanayo ephakathi) kanye nezingozi zokulahlekelwa ukulawula lapho amandla aphezulu ehlanganiswa nokuzimela okubanzi.
Ukucaca, ukuhunyushwa, kanye namahhashi eTrojan kumamodeli
Enye yezinselele ezinkulu ukuthi izinhlelo eziningi ze-AI zanamuhla, ikakhulukazi amanethiwekhi amakhulu e-neural, zisebenza njengamabhokisi amnyama angempela . Siyazi ukuthi yini engena kuzo nokuthi yini ephumayo, kodwa ukuqonda ukuthi kungani zenze isinqumo esithile, noma ukuthi yimaphi amaphethini angaphakathi abawasebenzisayo empeleni, akulona neze into encane.
Lokhu kuntuleka kokubonisa izinto obala kungqubuzana ngqo nezidingo zomthetho nezokuziphatha ukuze kuchazwe ezindaweni ezifana nesikweletu, ukukhethwa kwabasebenzi, kanye nohlelo lobulungiswa, lapho kuphoqelekile khona ukucacisa ukuthi kungani umuntu othile amukelwa noma enqatshwa. Kuphinde kube nzima ukuphenya ukwehluleka okukhulu: ezigamekweni ezifana nokushayisana okubulalayo nezimoto ezizimele, ngezinye izikhathi kunzima kakhulu ukwakha kabusha ukuthi yikuphi ukuhlanganiswa kwezimpawu zangaphakathi okuholele ephutheni ekuqondeni noma esinqumweni.
Yingakho kuthuthukiswe insimu yonke yokuhumusha kanye namasu okucaca . Abanye bagxila ekuboniseni ukuthi yiziphi izingxenye zokufaka amamodeli "azibukayo" (isibonelo, ukuthi yimaphi amaphikseli noma amagama abe nomthelela omkhulu ekukhishweni), kanti abanye baya ngokujulile, behlaziya ukumelwa kwangaphakathi kanye nokuxhumana phakathi kwama-neurons. Kumamodeli afana ne-CLIP, ama-neurons atholakale aphendula kokubili igama elithi "spider" kanye nezithombe zikaSpiderman, kanye nama-transformer olimi, kufundwa amaphethini okunaka kanye nezifunda eziphindaphindayo ezibonakala zifaka imithetho yokwenziwa noma ukuhlanganiswa kwamaqiniso.
Ngaphezu kokusiza ekuqondeni nasekulungiseni amaphutha, lawa masu avumela ukuhlela imodeli okuqondiwe . Kuye kwenzeka ukuthola amapharamitha ahlobene nomcimbi othize (isibonelo, idolobha lapho kukhona khona isikhumbuzo esidumile) bese uwaguqula ukuze ushintshe izimpendulo ngokuhlelekile. Nakuba lokhu kuhlola kubonisa ukuthi amaphutha angadalwa futhi, indlela efanayo ingasetshenziswa ukulungisa amaphutha amakhulu ngaphandle kokudinga ukuqeqesha kabusha amamodeli kusukela ekuqaleni.
Enye inkinga ebucayi amaTrojan noma ama-backdoor kumamodeli e-AI . Lokhu kubhekisela ezimweni lapho, ngesikhathi sokuqeqeshwa, othile ethula iphethini ethile, lapho ivela ekufakweni, ibangela ukuziphatha okuhlukile kunokujwayelekile. Isibonelo, uhlelo lokubona ubuso oluvumela kuphela ukufinyelela uma umuntu egqoke isesekeli esithile, noma imoto ezishayela yona eziphatha ngokunganaki uma ithola isitikha esithile esibhengezweni.
Isici esikhathazayo ukuthi ukushintsha ngisho nengxenyana encane yedatha yokuqeqesha kwanele ukufaka iTrojan esebenzayo. Njengoba amamodeli amaningi eqeqeshwe ngemininingwane eminingi yomphakathi, kunengqondo ukucabangela izimo lapho umhlaseli ekwazi ukungcolisa ingxenye ye-corpus. Lawa maTrojan, ngaphandle kokuba sengozini ngokwawo, asebenza njengendawo efanele yokuhlola yokuthuthukisa amathuluzi angcono okuhlola nokuthola ukuziphatha okubi okuyimfihlo.
Izingozi zesakhiwo, ukuncintisana kanye nokuphathwa komhlaba wonke
Ngaphandle kokuhlaselwa okuthile noma izehlakalo ezizimele, kunezingozi ezihlelekile ezihlobene nendlela uhlelo lwe-AI oluhlelwe ngayo: ingcindezi yokuncintisana phakathi kwezinkampani namazwe, ukuntuleka kwesiko lokuphepha, izisusa zomnotho zokukhulula imikhiqizo ngokushesha okukhulu, kanye nokungalingani phakathi kwabahlaseli be-cyber kanye nabavikeli.
Abahlaziyi abaningana baveze ukuthi ukugxila kuphela “ekuhlulekeni kobuchwepheshe” noma “ekusetshenzisweni kabi komuntu ngamunye” kuyasilela, ngoba izintambo zezimbangela eziholela enhlekeleleni zinde kakhulu . Ukuguquguquka kwesimo sezwe ngesikhathi seMpi Yomshoshaphansi—lapho izinqumo zabantu abambalwa zingasondeza khona noma ziqhelelane nayo—zivame ukubizwa njengesifaniso salokho okungenzeka ngomjaho ophithizelayo wokuthuthukisa izinhlelo ze-AI ezinamandla ngokwengeziwe.
Ukwesaba okuphindaphindiwe yilokho okubizwa ngokuthi “umjaho oya kwalasha” : uma abadlali abazimele noma bakahulumeni bebona izinzuzo ezibalulekile zokuba phambili (kwezomnotho, kwezempi, noma ngokwesikhundla), bangase bazizwe bephoqelekile ukusheshisa nokunciphisa izilawuli zokuphepha, noma ngabe wonke umuntu uyavuma ukuthi intuthuko ehamba kancane ingaba ngcono. Ukugwema lokhu kuguquguquka kudinga izindlela zokubambisana, ukucaca okuhlangene, izivumelwano ngezindinganiso eziphansi, kanye, cishe, nezikhungo ezintsha zomhlaba wonke.
Ngesikhathi esifanayo, izinselelo eziningi zomhlaba wonke zanamuhla ziyizinkinga zokubambisana : ukuguquka kwesimo sezulu, izingozi zenuzi, izikhali ze-cyber, njalo njalo. I-AI ingenza lezi zinselelo zibe zimbi kakhulu (isibonelo, ngokwenza kube lula ukuhlaselwa kwe-cyber okunamandla kakhulu noma imikhankaso yokuphatha eyinkimbinkimbi kakhulu yezepolitiki), kodwa ingasiza futhi ngokunikeza amathuluzi angcono okubikezela, ukuhlaziywa kwezimo, kanye nokwenza izinqumo ezihlangene.
Ngakho-ke ukuhlanganiswa kwensimu yokuphathwa kwe-AI , efuna ukuchaza izinhlaka zomthetho, amazinga obuchwepheshe, imikhuba emihle yebhizinisi, kanye nezindlela zokuziphendulela. Kusukela emisebenzini eyisisekelo eqhathanisa umthelela ongaba khona we-AI nobuchwepheshe obushintshayo besikhathi esidlule (njengogesi noma injini yomusi) kuya eziphakamisweni eziqondile kakhulu: ukuhlolwa kwangaphandle , izinzuzo zamaphutha, izizindalwazi zezehlakalo zomphakathi, izindlela zokunquma ukuthi nini lapho kunesibopho sokushicilela imiphumela noma amamodeli, kanye nezidingo zokuphepha ezithuthukisiwe zama-laboratory aholayo.
Indima yokuphepha kwe-cyber: I-AI yokuvikela nokuvikela i-AI uqobo lwayo
Emsebenzini wansuku zonke, i-AI isivele ingumdlali oyinhloko kwezokuphepha kwe-inthanethi , futhi ubudlelwano buhamba ngazo zombili izindlela: sisebenzisa i-AI ukuze sizivikele kangcono, futhi ngesikhathi esifanayo kufanele sivikele izinhlelo ze-AI ngokwazo ekuhlaselweni.
Emashumini eminyaka amuva nje, izinsongo ze-cyber ziye zashintsha kusukela ekufuneni ukuphazamiseka okubonakalayo kuya ekwebeni idatha, ukungena ezinhlelweni izinyanga eziningi, kanye nokusebenzisa i-malware eyimfihlo kakhulu . Ukuze kulwiwe nazo, izixazululo ezifana ne-SIEM, eqoqa futhi ihlaziye inani elikhulu lamalogi, kanye ne-EDR, eqapha njalo okwenzekayo kuwo wonke amadivayisi enkampani, ziye zathandwa. Inkinga ukuthi inani ledatha lidlula kakhulu lokho iqembu labantu elingakubukeza ngesandla.
Yilapho ukufunda komshini kungena khona, okukwazi ukubona amaphethini kanye nokungafani ngesikhathi sangempela kusuka kumasethi amakhulu ezimpawu: imizamo yokungena ngemvume engavamile, ukufinyelela okungagunyaziwe kudatha ebucayi, ukuziphatha kwenqubo okungajwayelekile, noma ukuxhumana nezizinda ezisolisayo.
Amapulatifomu e-AI kwezokuphepha nawo asetshenziselwa ukwenza ingxenye yempendulo isebenze ngokuzenzakalela : ukuhlanganisa izexwayiso ezihlobene, ukwakha kabusha "indaba" yokuhlaselwa, ukubeka phambili ukuthi yiziphi izehlakalo okufanele zibukezwe kuqala, futhi, kwezinye izimo, ukwenza ngqo izenzo zokuvimbela (ukuhlukanisa umshini, ukusika uxhumano, ukususa iziqinisekiso ezithintekile). Lokhu kunciphisa umsindo, kusiza abahlaziyi ukugxila kulokho okubalulekile, futhi kunciphisa kakhulu isikhathi kusukela lapho inkinga itholakala kuze kube yilapho kuthathwa isinyathelo esisebenzayo.
Ngesikhathi esifanayo, intuthuko efanayo iyasetshenziswa ekuvikelweni kwamanethiwekhi kanye nama-endpoints . Ama-firewall esizukulwane esilandelayo ahlanganisa amamodeli ahlaziya ithrafikhi ukuze athole ukuziphatha okunonya, abuyekeze imithetho yangaphakathi ngokushesha, noma ahlele kabusha ukuhlukaniswa kwenethiwekhi ukuze anciphise ukunyakaza kwabahlaseli eceleni. Ama-EDR athuthukisiwe yi-AI athola ukuziphatha kwenqubo okungavamile (njengokufinyelela okungalindelekile kumafayela abucayi noma uchungechunge lokusebenzisa olujwayelekile lwe-ransomware) ngaphambi kokuba umonakalo ungalungiseki.
Okunye okungeziwe kukho konke lokhu ukwanda kwabasizi abasekelwe olimini lwemvelo abahlanganiswe namathuluzi okuphepha kwe-inthanethi. Laba bashayeli be-AI abasiza amaqembu okuphepha babavumela ukuthi babuze imibuzo eyinkimbinkimbi ("Yiziphi izindlela zokuhlasela ezikhona kule database?") futhi bathole izincazelo ezihambisana nomongo, izifinyezo zezehlakalo, noma iziphakamiso zokunciphisa ngaphandle kokuzulazula ngesandla ngamadeshibhodi amaningi.
Indlela abahlaseli abasebenzisa ngayo i-AI
Ngeshwa, izigebengu ze-inthanethi nazo ziye zangena ngenkani ekusebenziseni i-AI . Ezinkundleni eziyimfihlo, yonke imodeli entsha enamandla evelayo iyahlaziywa, futhi kuvele izinguqulo eziguquliwe, ezinye ziqondiswe ngqo emisebenzini yobugebengu, zasuswa izihlungi zokuziphatha futhi zaqeqeshwa "njengabasizi bokugenca."
Namuhla kulula kakhulu ukusizakala ngamamodeli akhiqizayo ukuze zenzekalise cishe zonke izigaba zomsebenzi wobugebengu : ukubhala ngobuningi ama-imeyili e-phishing angenaphutha futhi enziwe ngokwezifiso, ukudala amakhasi okungena mbumbulu athembekile kakhulu, ukukhiqizwa kwezikripthi ze-ransomware noma amakhithi okweba iziqinisekiso , ukucaciswa kwemiyalezo ekholisayo yobunjiniyela bezenhlalo, njll.
Enye into ekhathazayo yikhono le-AI lokudala okuqukethwe okungelona iqiniso okukholisayo . Amathuluzi okukhiqiza umbhalo avumela ukukhiqizwa kwezindatshana ezingamanga nezindaba zezindaba ezilingisa isitayela semidiya yobungcweti; izinhlelo zokukhiqiza amavidiyo njenge-deepfakes zithatha indawo yobuso ngeziqeshana zekhwalithi ephezulu; futhi ukwenziwa kwezwi ngendlela efanayo kwenza kube nokwenzeka ukuphinda izwi lomuntu ngemizuzu embalwa yomsindo.
Amacala okukhwabanisa ngocingo kusetshenziswa amazwi aqoshwe amalungu omndeni ukuze kukhuthazwe ukudluliselwa kwemali asevele ebhaliwe , kanye nemikhankaso yokunyundela yezepolitiki kanye nokukhwabanisa ezinkundleni zokuxhumana okuhlanganisa amavidiyo aguquliwe, umsindo mbumbulu, kanye nama-robot akhiqizayo ukuze kuthuthukiswe imiyalezo. Izindleko zala mathuluzi ziyancipha minyaka yonke, okwandisa imbuyiselo yokutshalwa kwezimali kwabahlaseli.
Ngamafuphi, i-AI yandisa amakhono okuhlasela: yenza ukuhlaselwa okuyinkimbinkimbi kube lula futhi kushibhile, ivumela imikhankaso emikhulu, futhi yenza kube nzima umsebenzi wokuhlukanisa ulwazi oluyiqiniso nolushintshiwe . Ngakho-ke, ukusebenzisa i-AI ngokuzivikela akusadingeki: ngaphandle kokuzenzakalela kanye nobuhlakani, amaqembu abantu awakwazi ukuqhubeka.
Izingozi ezibalulekile zokuphepha ze-AI ethile
Ngaphandle kwezingozi zokuphepha kwe-inthanethi ezijwayelekile, ukusebenzisa izinhlelo ze-AI kuhilela ubuthakathaka bazo okumele buqondwe futhi buphathwe. Ezinye zezinto ezibaluleke kakhulu yilezi:
Idatha kanye noshevu wemodeli. Ukufaka idatha enonya noma eguquliwe kumasethi okuqeqesha kungaphambukisa ukuziphatha kwemodeli, kufake iminyango yangemuva, noma kwehlise ukusebenza ngaphandle kokubonakala ezilinganisweni ezijwayelekile. Ezindaweni zokufunda ezihlanganisiwe, lapho abadlali abaningi befaka isandla ekuqeqeshweni, ukuthola ukuthi yimuphi umhlanganyeli ofake umsindo ngamabomu kuyinselele kakhulu.
Izibonelo eziphikisanayo kanye nokugwema. Ukuklama okokufaka okushintshiwe kancane ukuze kubangele amaphutha ahlelekile kuvumela abahlaseli ukuthi badlule izinhlelo zokuthola i-malware, izihlungi zogaxekile, izihlungi zezithombe, noma izindlela zokuqinisekisa i-biometric. Ngenxa yokuthi imodeli "ikholelwa" ukuthi okufakwayo kuyinto evamile, lezi zinhlobo zokuhlasela kunzima ukuzinciphisa kusetshenziswa imithetho yendabuko kuphela.
Ukuguqulwa kwemodeli kanye nokuvuza kobumfihlo. Ngemibuzo ephindaphindwayo kanye nokuhlaziywa komphumela, umhlaseli angakha kabusha ulwazi mayelana nedatha yokuqeqesha, aze aphethe ngokuthi ngabe irekhodi elithile lifakiwe noma akhiphe izingcezu zedatha ebucayi egcinwe yimodeli. Lokhu kudala inkinga enkulu lapho kuqeqeshwa ngolwazi lomuntu siqu, lwezokwelapha, noma oluyimfihlo.
Ukwebiwa kwamamodeli kanye nempahla yobuhlakani. Ngokubuza kakhulu i-AI API, kungenzeka ukuqeqesha imodeli yokukopisha ephinda ngokuseduze ukuziphatha kokuqala. Lokhu kwehlisa ikhwalithi yebhizinisi yalabo abanikeza i-AI njengesevisi futhi, ngaphezu kwalokho, kudala ikhophi lapho umhlaseli elawula khona ngokuphelele, okuhlanganisa nekhono lokusebenzisa ubuthakathaka ngaphandle kwemingcele.
Ukuhlaselwa kwama-API kanye nengqalasizinda. Izixhumi ezixhumanisa amamodeli e-AI nezinye izinhlelo zokusebenza ziyindawo yokungena ehehayo kakhulu: ukuqinisekiswa okubuthakathaka, ukuntuleka kokuqinisekiswa kokufaka, ukungabikho kwemingcele yemvamisa noma ukuqapha kungavumela noma yini kusukela ekwenqatshweni kwensizakalo kuya ekukhishweni kwedatha okukhulu noma ekusetshenzisweni kwekhodi ekude.
Ukujova okusheshayo kanye nokusetshenziswa kabi ku-GenAI. Endabeni yamamodeli olimi kanye nabasizi bokukhiqiza, i-vector evelayo yileyo yokujova kwemiyalelo eqondile noma engaqondile . Umhlaseli angashumeka imiyalo enonya embhalweni wokufaka uqobo ("angayinaki imithetho yangaphambilini kanye…") noma kumadokhumenti angaphandle azofundwa yimodeli, ngenhloso yokudlula imikhawulo, ukukhipha idatha, noma ukwenza izenzo ezingahlosiwe.
Ukuxhaphaza imibono engekho. Amamodeli akhiqizayo angakha idatha ngokukhululeka okuphelele. Uma lezi zinhlelo zisetshenziswa ngaphandle kokubhekwa ngumuntu ezimweni ezibucayi, noma uma abahlaseli bekwazi ukwenza imibono ethile engekho “njengamaqiniso,” kuvula umnyango wezinqumo eziyiphutha, ulwazi olungamanga, kanye nokulahlekelwa ukwethenjwa kolwazi lwedijithali okusabalele.
Izinhlaka, amazinga, kanye nemikhuba emihle kakhulu ye-AI ephephile
Ukuze kugwenywe ukungena endaweni engaboni, izinhlangano ezahlukahlukene ziphakamise izinhlaka ezithile zokuphepha kwe-AI . Nakuba kungekho ezimboza yonke indawo ngokwazo, zizonke zinikeza isisekelo esiqinile sokwakha isu.
Phakathi kwazo, uHlaka Lokuphathwa Kwengozi lwe-NIST AI luvelele , luhlela umsebenzi ozungeze imisebenzi emine: ukubusa (ukusungula izinqubomgomo kanye nemithwalo yemfanelo), ukumaka (ukuqonda umongo kanye nezingozi), ukukala (ukuhlola ukusebenza kanye nokuvezwa), kanye nokuphatha (ukubeka phambili nokusebenzisa izilawuli). Eminye imizamo, njenge-MITRE ATLAS matrices, ikhathalogi yamaqhinga athile okuhlasela kanye namasu ngokumelene nezinhlelo ze-AI, okufana nalokho okwenziwa yi-ATT&CK ngezinsongo zendabuko.
Ensimini esetshenziswa kakhulu, umphakathi wezokuphepha usungule uhlu olufana ne- OWASP Top 10 ye-LLM , ehlanganisa izingozi ezinkulu kuzinhlelo zokusebenza ngokusekelwe kumamodeli ezilimi ezinkulu: ukufakwa okusheshayo, ukudalulwa kwedatha, ubuthakathaka ochungechungeni lokuhlinzekwa kwamamodeli, ukwebiwa kwamamodeli, njll. Abathengisi abakhulu bashicilele izinhlaka zabo (njenge-SAIF ye-Google) ezinezinyathelo eziqondile zokusebenzisa izinhlelo ze-AI ngokuphephile kakhudlwana.
Ngaphezu kwalokho, kuvela izindlela eziqondile kakhulu nezisebenzayo zokusebenzisa i-GenAI ezindaweni zamafu, njengezinhlaka zokuhlukanisa abaqashi kanye nabasebenzisi ezihlukanisa ngokucophelela idatha, izimvume, kanye neziteshi zokuxhumana, ukunciphisa umthelela wokwehluleka noma ukufakwa okungenzeka. Umqondo ukuhlanganisa izilawuli zokuphepha kwe-cyber zendabuko (ukubethela, ukuphathwa kobunikazi, ukuhlukaniswa kwenethiwekhi, i-sandboxing) nezinyathelo ezenzelwe izici ezithile ze-AI.
Ngokusekelwe kulokhu, izincomo eziphindaphindwayo kakhulu zifaka phakathi ukuqinisekisa ikhwalithi kanye nokuhlukahluka kwedatha yokuqeqesha, ukubhekana ngqo nemiphumela yokuziphatha (ukubandlulula, imvume, ubumfihlo), ukwenza ukuhlolwa kwemodeli njalo kanye nokuvuselelwa, kanye nokuchaza izinqubomgomo zokuphepha ze-AI ezihambisana nezinqubomgomo ze-ICT ezijwayelekile.
Ngesikhathi esifanayo, kuvela isigaba sezixazululo esaziwa ngokuthi i-AI Security Posture Management (AISPM) , esifana nalokho i-CSPM eyayikusho ngefu: ukuqopha izimpahla ze-AI, ukuthola ukusetshenziswa okungagunyaziwe ("isithunzi se-AI"), ukuhlola ukucushwa, kanye nokuqapha njalo ukuhambisana nenqubomgomo.
Konke kusikisela ukuthi, njengoba i-AI ihlanganiswa kakhulu nengqalasizinda ebalulekile, ukuphepha kwe-AI kuzoshintsha kusuka "kokungeziwe okuthakazelisayo" kuye kwesidingo esiyisisekelo . Izinhlangano ezifika phambili—ngokugcina izinhlelo zazo ze-AI, ukuchaza ukubusa okucacile, ukusebenzisa izilawuli zokufinyelela eziqinile, ukufaka amamodeli azo ekuhlolweni okuphikisayo, nokuzivumelanisa nezinhlaka eziqashelwayo—zizoba sesimweni esingcono kakhulu sokusebenzisa izinzuzo ze-AI ngaphandle kokubhekana nezimanga ezingathandeki. Futhi, ngaphezu kwakho konke, zizohlonyiswa kangcono ukugcina ukwethenjwa kwamakhasimende, abasebenzisi, kanye nezakhamizi endaweni lapho ukuhlukanisa phakathi kwalokho okungokoqobo nalokho okuguquliwe kuzoba nzima kakhulu.
