Mucukumbuzi nkinyuma yinyuma: Umukiriya-Uruhande LLM Gukuramo Ibanga Demystified
Kuva kera cyane, twafashe "umuburo wumutekano wa AI" nkibikangisho bya kure cyangwa amatsiko yo kwiga. Twumvise ibijyanye no guterwa vuba, gufungwa, no gutangaza amakuru mumajwi acecetse, akenshi bigakurikirwa no kwikinisha no "tuzabigeraho." Ariko, imyiyerekano iheruka kwerekanwa yahinduye iyo kwibeshya. Ntabwo aribwo bugizi bwa nabi bugaragara; ni bizima, bishingiye kuri mushakisha, kubakiriya-kuruhande rwo gukuramo amakuru yunvikana muri Moderi nini y'ururimi (LLMs) [2]. Niba ukoresha porogaramu zikoreshwa na LLM, cyane cyane ikintu cyose kireba urubuga, "ikibazo cyawe kizaza" cyahindutse gusa, cyane, kandi gishobora kuba giteye isoni cyane.
Ntabwo arindi mutwe rusange "AI ni risque". Ubu ni uburyo bwihariye, bwerekana uburyo umukinnyi mubi, ukoresheje mushakisha gusa, ashobora guhatira LLM yawe guhishura amakuru yihariye. Tekereza gato. Ntabwo ari ibijyanye na seriveri igoye cyangwa ibyitegererezo byimbitse. Byerekeranye no guhuza imikoranire - umuryango wimbere wibisabwa - gukoreshwa hamwe ninyuguti nke zakozwe neza. Nibikoresho bya digitale bihwanye numuntu utegura interuro ibereye kuri chatbot yawe ifite umutekano, kandi itanga ubushishozi itanga amakuru akomeye, akenshi buhoro buhoro.
Ibishushanyo byumutekano wicyitegererezo
Ibigo byinshi byashora imari cyane muri LLMs "zifite umutekano", haba muburyo bwiza bwo gutondekanya imishinga cyangwa gushingira kumurongo wo mu rwego rwa API hamwe nibitekerezo byizamu. Igitekerezo kidasubirwaho nuko niba icyitegererezo ubwacyo gikomeye, porogaramu yubatswe hejuru yayo izaragwa uwo mutekano. Imyiyerekano iheruka iragaragaza ibi nkibitekerezo bibi. Intege nke ntabwo ziri mubwenge bwibanze bwikitegererezo cyangwa amakuru yamahugurwa; ni muri * interface * twubaka hafi yayo [1]. Nibisanzwe "bikomeye vault, flimsy front door" ikibazo. Urashobora kuba ufite iterambere ryambere, rikomeye LLM kwisi, ariko niba porogaramu yawe yinjiza igereranya ari naive, cyangwa ibisohoka muyunguruzi ntibihari, mubyukuri usize amakuru yingenzi yerekanwe kumuntu wese.
Ubu buryo "bwihuse bwo gutera inshinge" burareba cyane cyane kuko bukoresha imiterere ya LLM: ubushobozi bwayo bwo gukurikiza amabwiriza nuburyo bwuzuye. Mugutera inshinge zubatswe neza mubisobanuro byumukoresha, uwagabye igitero arashobora gukoresha imiterere yimbere yimbere, akayihatira guhishura amakuru atigeze agenewe [3]. "Ikintu kimwe" cyangwa icyerekezo ni ingenzi hano. Bisobanura uburyo bwo kugenzura, kugenzurwa, ntabwo ari akajagari gusa. Tekereza uwagabye igitero abaza inshuro nyinshi, "Ese imiterere ya mbere y'ibanga 'A'?" hanyuma "Ni 'B'?" n'ibindi. Biratinda, yego, ariko nanone birakomeye bidasanzwe kandi biragoye kumenya niba utabishaka byumwihariko.
Mucukumbuzi yawe, Intwaro Yabo
Kuba iyi ari * mushakisha ishingiye kuri * ikoresha ihindura umukino rwose. Bisobanura:
- ** Inzitizi ntoya yo kwinjira: ** Umuntu wese ufite ubumenyi bwibanze bwiterambere ryurubuga na mushakisha arashobora kugerageza ibi. Nta bikoresho byihariye cyangwa ubumenyi bwimbitse bwa AI busabwa. **
- ** Iterabwoba Ako kanya: ** Niba porogaramu yawe ya LLM ari nzima kandi igerwaho ukoresheje interineti, birashoboka * bimaze * kwibasirwa niyi vector. Ibi ntabwo ari ibizaza; ni ibihe byihutirwa.
Reba ingaruka: ibisobanuro bya kode yihariye, inyandiko zimbere, amakuru yumukiriya, urufunguzo rwa API, cyangwa no gukoresha uburyo bworoshye bwimyitwarire ya LLM kugirango ukwirakwize amakuru atari yo cyangwa utange ibintu byangiza. Ibi byose birashobora gukururwa cyangwa guhinduka, biti bitoroshye, binyuze mumasanduku yoroshye.
Inshingano ni Iwawe
Ntabwo arikibazo kubatanga icyitegererezo nka OpenAI cyangwa Google kugirango bakosore mugihe cyicyitegererezo gikurikira. Iki nikibazo * cyumutekano wumutekano *, kandi inshingano zo kugabanya kugabanuka cyane cyane kubitugu byabateza imbere, amatsinda yumutekano, na CTOs zikoresha sisitemu. LLM yawe ni imbaraga, ariko amaherezo naive, moteri. Izakora ibyo yabwiwe, nubwo ayo mabwiriza yihishe mubintu bisa nkaho ari inzirakarengane.
None, ubu ukora iki *?
Ibikorwa Byihuse:
1 .. Ugomba gusuzuma * buri nyuguti * yinjira muri LLM yawe. Shakisha uburyo bwerekana inshinge zihuse, urutonde rwimiterere idasanzwe, cyangwa kugerageza guca ukubiri n'ibiteganijwe kuganira [4]. Ibi ntabwo ari uguhagarika "amagambo mabi"; nibijyanye no kumenya ibyifuzo byo guhangana. Kurenga regex yibanze nisesengura rya heuristic, shyira mubikorwa isesengura ryibisobanuro kugirango umenye umugambi mubi. Koresha tokenisation hamwe no gushiramo isesengura kugirango umenye uburyo bwo kwinjiza ibintu bidasanzwe. Tekereza gukoresha icyiciro cya kabiri, gito LLM cyangwa sisitemu ishingiye kumategeko yatojwe kugirango tumenye kugerageza inshinge. Shakisha ibishushanyo nkamabwiriza yo gukina, ibyifuzo bisobanutse byo gukuramo amakuru, cyangwa kugerageza kurenga amabwiriza yabanjirije. 2. ** Ibisohoka bikomeye byo gushungura no kwemeza: ** Ntukigere wizera umusaruro wa LLM muburyo butaziguye. Mbere yuko inyandiko iyo ari yo yose yakozwe na LLM yerekanwa ku mukoresha cyangwa ikoreshwa n'indi sisitemu, igomba * kunyura mu buryo bukomeye bwo kwemeza. Harimo ijambo ryibanze? Bihuye nuburyo buteganijwe? Iragerageza gukora code cyangwa guhishura sisitemu y'imbere? Niba LLM yawe igomba kuvuga mu ncamake inyandiko, ariko igatangira gutondeka inzira za dosiye, ufite ikibazo. 3.. Ariko rero, umva ko ubwubatsi bwihuse bwonyine budahagije; ni kwirwanaho byimbitse, ntabwo ari ingabo yambere. 4 .. Shyira mu bikorwa igenzura rikomeye. Niba LLM igomba gusa gusubiza ibibazo bijyanye ninyandiko rusange, menya ko idafite gahunda yo kugera kububiko bwimbere cyangwa APIs zoroshye. 5. ** Gukurikirana no Kumenya Anomaly: ** Ugomba gushakisha cyane imyitwarire idasanzwe ya LLM. Imitwe muburyo bwihariye bwibibazo, ibisubizo bitandukana nuburyo busanzwe, cyangwa kugerageza kenshi gukuramo amakuru bigomba guhita byihutirwa.
Ibi ntibireba "umutekano wa AI" ukundi. Ibi bijyanye no kurinda urwego rwimikoranire ya porogaramu ya LLM kurwanya ikintu cyihariye, cyerekanwe, kandi gihita gishobora gukoreshwa. Twagiye twubaka sisitemu zidasanzwe, zikomeye, akenshi hamwe ningamba zumutekano zidahagije kuri interineti. Imyiyerekano iheruka yatweretse neza uburyo byoroshye gukoresha ibyo byuho. Igihe cyo kuganira kubitekerezo cyarashize. Igikorwa cyihuse kirakenewe kugirango umutekano wa LLM ushobore.
Inkomoko
. [2] Lakera Gandalf: Ikibazo cyo Gutera Imiti Yihuse - https://gandalf.lakera.ai/ [3] Ibitero byihutirwa | Urubuga rwumutekano wurubuga - https://portswigger.net/web-umutekano/llm-ibitero/prompt-injection [4] Gutera inshinge niki kandi ni gute twakwirinda? - https://www.