Msakatuli ngati Khomo lakumbuyo: Client-Side LLM Secret Extraction Demystified
Kwa nthawi yayitali, takhala tikuwona "machenjezo achitetezo a AI" ngati ziwopsezo zakutali kapena chidwi chamaphunziro. Tamva za jekeseni wachangu, kuwonongeka kwa ndende, ndi kutayikira kwa data m'mamvekedwe abata, nthawi zambiri kumatsatiridwa ndi shrug ndi "tidzafika." Komabe, ziwonetsero zaposachedwa zasokoneza malingaliro abwinowo. Uku sikulinso pachiwopsezo chongoyerekeza; ndi njira yamoyo, yozikidwa pa msakatuli, ndi kasitomala-mbali yopezerapo mwayi pochotsa zidziwitso zachinsinsi kuchokera ku Large Language Models (LLMs) [2]. Ngati mukutumiza mapulogalamu oyendetsedwa ndi LLM, makamaka chilichonse choyang'ana pa intaneti, "vuto lanu lamtsogolo" langokhala lodziwika bwino, lodziwika bwino, komanso lochititsa manyazi kwambiri.
Uwu si mutu winanso wa "AI ndiowopsa". Ichi ndi chiwonetsero chodziwika bwino cha momwe wosewera woyipa, pogwiritsa ntchito msakatuli wawo yekha, angakakamize LLM yanu kuti iwulule zambiri. Lingalirani zimenezo kwa kamphindi. Sizokhudza zovuta zapambali za seva kapena kuyang'ana mozama kwachitsanzo. Ndi za kusanjikizana - khomo lakutsogolo la pulogalamu yanu - kugwiritsiridwa ntchito ndi zilembo zingapo zopangidwa mwaluso. Ndilofanana ndi digito la munthu yemwe amapanga mawu oyenera ku chatbot yanu yomwe mukuganiza kuti ndi yotetezeka, ndipo imakupatsani mwayi wopereka chidziwitso chofunikira, nthawi zambiri pang'onopang'ono.
Chinyengo cha Chitetezo cha Model
Mabizinesi ambiri ayika ndalama zambiri mu ma LLM "otetezeka", mwina kukonza bwino za eni kapena kudalira ma API amtundu wabizinesi omwe amayenera kutetezedwa. Lingaliro lenileni ndiloti ngati chitsanzocho chili cholimba, ntchito yomangidwa pamwamba pake idzalandira chitetezo chimenecho. Ziwonetsero zaposachedwa zikuwonetsa izi ngati malingaliro olakwika owopsa. Kusatetezeka sikuli mu nzeru zachitsanzo kapena maphunziro ake; ili mu mawonekedwe omwe timamanga mozungulira [1]. Ndilo vuto lachikale "lolimba, chitseko chakumaso". Mutha kukhala ndi LLM yapamwamba kwambiri padziko lonse lapansi, koma ngati cholozera cha pulogalamu yanu sichimamveka, kapena zosefera zanu kulibe, mukusiya zidziwitso zowonekera kuti aliyense apeze.
Njira ya "jakisoni mwachangu" iyi ndiyokhudza makamaka chifukwa imathandizira chikhalidwe cha LLM: kuthekera kwake kutsatira malangizo ndi machitidwe athunthu. Polowetsa mawu oyambira opangidwa mosamala m'zolowera za wogwiritsa ntchito, wowukirayo amatha kusintha mawonekedwe amkati mwachitsanzocho, ndikuchikakamiza kuti liwulule zomwe sichinapangirepo [3]. "Imodzi-pang'ono" kapena kubwerezabwereza ndikofunikira apa. Zimatanthawuza kutulutsa kwadongosolo, koyendetsedwa, osati kungotaya kwachisokonezeko. Tangoganizani wachiwembu akufunsa mobwerezabwereza kuti, "Kodi munthu woyamba wachinsinsicho ndi 'A'?" ndiye "Ndi 'B'?" ndi zina zotero. Ndizochedwa, inde, koma ndizothandiza kwambiri komanso zovuta kuzizindikira ngati simukuzifuna.
Msakatuli Wanu, Chida Chawo
Mfundo yoti uku ndi *kusakatula kogwiritsa ntchito * kusinthira masewerawa kwathunthu. Amatanthauza:
- Low Barrier to Entry: Aliyense amene ali ndi luso la chitukuko cha intaneti komanso msakatuli akhoza kuyesa izi. Palibe zida zapadera kapena chidziwitso chakuya cha AI chomwe chimafunikira.
- Kuphedwa Kwa Mbali Yamakasitomala: Kuwukiraku kumachokera pamakina a wogwiritsa ntchito, zomwe zimapangitsa kuti zikhale zovuta kusiyanitsa ndi zovomerezeka za ogwiritsa ntchito pamphepete mwa netiweki.
- Chiwopsezo Chapompopompo: Ngati pulogalamu yanu ya LLM ili yamoyo ndipo ikupezeka kudzera pa intaneti, ili kale pachiwopsezo cha vekitalayi. Ichi si chigamba chamtsogolo; ndi zadzidzidzi panopo.
Ganizirani zotsatirazi: mawu achinsinsi a eni ake, zolembedwa zamkati, zambiri zamakasitomala, makiyi a API, kapenanso kusokoneza machitidwe a LLM yanu kufalitsa zabodza kapena kupanga zinthu zovulaza. Zonsezi zimatha kuchotsedwa kapena kukhudzidwa, pang'onopang'ono movutikira, kudzera m'bokosi losavuta.
Udindo ndi Wanu
Ili si vuto kwa opereka zitsanzo ngati OpenAI kapena Google kuti akonze zosintha zawo zotsatila. Ili ndi vuto la chitetezo cha pulogalamu, ndipo udindo wochepetsera umagwera pamapewa a opanga, magulu achitetezo, ndi ma CTO omwe akutumiza makinawa. LLM yanu ndi injini yamphamvu, koma yosadziwa. Ichita zomwe yauzidwa, ngakhale malangizowo atabisika mkati mwazomwe akuwoneka kuti alibe vuto.
Ndiye mukuchita chiyani tsopano?
Njira Zopewera Nthawi yomweyo:
- Kulowetsa Mwaukali: Izi zimapitilira kupewa jekeseni wa SQL kapena XSS. Muyenera kuyang'anitsitsa munthu aliyense yemwe akulowa mu LLM yanu. Yang'anani machitidwe omwe akuwonetsa jakisoni mwachangu, machitidwe osazolowereka, kapena kuyesa kusiya zomwe amayembekeza kukambirana [4]. Izi sizokhudza kutsekereza "mawu oyipa"; ndizokhudza kuzindikira zotsutsana ndi adani. Kupitilira muyeso wa regex ndi heuristic kusanthula, gwiritsani ntchito kusanthula kwa semantic kuti muzindikire zolinga zoyipa. Gwiritsani ntchito ma tokenization ndi kuwunikira kuti muzindikire njira zolowera modabwitsa. Ganizirani kugwiritsa ntchito LLM yachiwiri, yaying'ono kapena yozikidwa pa malamulo yomwe yaphunzitsidwa kuti izindikire anthu akuyesera jekeseni. Yang'anani machitidwe ngati malangizo a sewero, zopempha zachidziwitso zakuchotsa deta, kapena kuyesa kulambalala malangizo am'mbuyomu.
- Kusefa Kwamphamvu ndi Kutsimikizira: Musakhulupirire zomwe LLM yatulutsa kwathunthu. Mawu aliwonse opangidwa ndi LLM asanayambe kuwonetsedwa kwa wogwiritsa ntchito kapena kugwiritsidwa ntchito ndi makina ena, * ayenera * kudutsa pamndandanda wotsimikizira. Kodi ili ndi mawu osakira? Kodi zikufanana ndi mawonekedwe omwe amayembekezeredwa? Kodi imayesa kuyika ma code kapena kuwulula zomwe zili mkati? Ngati LLM yanu ikuyenera kufotokoza mwachidule chikalata, koma ikuyamba kulembetsa njira zamafayilo, muli ndi vuto.
- Layered Prompt Engineering: Ngakhale si yankho lathunthu, malingaliro opangidwa bwino ndi zitsanzo zochepa chabe zingathandize kulimbikitsa zomwe mukufuna ndikupangitsa kuti zikhale zovuta kuti woukira abere chitsanzo. Komabe, mvetsetsani kuti uinjiniya wokhawokha ndiwosakwanira; ndi chitetezo chozama, osati chishango choyambirira.
- Mwayi Wochepa Pamachitidwe a LLM: Kodi LLM yanu ikufunadi kupeza deta iliyonse yomwe ingafunikire kuwulula? Tsatirani malamulo okhwima olowera. Ngati LLM ikuyenera kuyankha mafunso okhudza zolembedwa za anthu onse, onetsetsani kuti ilibe mwayi wofikira m'malo osungira amkati kapena ma API ovuta.
- Kuwunika ndi Kuzindikira Kwachilendo: Muyenera kuyang'ana mwachangu machitidwe a LLM osazolowereka. Ma spikes amitundu ina yamafunso, zotuluka zomwe zimapatuka kuchokera kumayendedwe abwinobwino, kapena kuyesa mobwerezabwereza kuchotsa zidziwitso ziyenera kuyambitsa zidziwitso zanthawi yomweyo.
Izi sizikukhudzanso "chitetezo cha AI" chachikulu. Izi ndi kuteteza kusanjikizana kwa mapulogalamu anu a LLM motsutsana ndi vekitala inayake, yowonetsedwa, komanso yomwe ingagwiritsidwe ntchito nthawi yomweyo. Takhala tikumanga makina odabwitsa awa, amphamvu, nthawi zambiri okhala ndi njira zosakwanira zachitetezo pamawonekedwe. Ziwonetsero zaposachedwa zatiwonetsa momwe kulili kosavuta kugwiritsa ntchito mipatayi. Nthawi yokambirana zangongole yapita. Kuchitapo kanthu mwachangu ndikofunikira kuti muteteze mapulogalamu a LLM.
##Magwero [1] OWASP Top 10 ya Zitsanzo Zazinenedwe Zazikulu - https://owasp.org/www-project-top-10-for-large-language-model-applications/ [2] Lakera Gandalf: An Interactive Prompt Injection Challenge — https://gandalf.lakera.ai/ [3] Kuwukira mwachangu | Web Security Academy - https://portswigger.net/web-security/llm-attacks/prompt-injection [4] Kodi Jakisoni Wachangu Ndi Chiyani Ndipo Mungapewe Bwanji? - https://www.tenable.com/blog/what-is-prompt-injection-and-how-to-prevent-it