Ukusetshenziswa kwabathuthukisi be-software kwamamodeli ezilimi ezinkulu (LLMs) kunikeza ithuba elikhulu kunalokho obekucatshangwa ngaphambilini ukuze abahlaseli basabalalise amaphakheji anonya ezindaweni zokuthuthukiswa, ngokocwaningo olusanda kukhishwa.

Ucwaningo oluvela kumthengisi wezokuphepha we-LLM i-Lasso Security luwukulandelela umbiko wangonyaka odlule mayelana nokungenzeka abahlaseli ukuhlukumeza ukuthambekela kwama-LLM ekuboneni izinto ezingekho, noma ukukhiqiza okubonakala kunesisekelo kodwa okungasekelwe eqinisweni, kuphumela ekuphenduleni kokufaka komsebenzisi.

I-AI Package Hallucination

The isifundo sangaphambilini igxile ekuthambekeni kwe-ChatGPT ukwenza amagama emitapo yolwazi yamakhodi - phakathi kwezinye izinto eziqanjiwe - lapho abathuthukisi be-software becela usizo lwe-chatbot enikwe amandla i-AI endaweni yokuthuthukiswa. Ngamanye amazwi, i-chatbot kwesinye isikhathi yayikhipha izixhumanisi zamaphakheji angekho kumakhosombe amakhodi omphakathi lapho unjiniyela engase ayicele ukuthi iphakamise amaphakheji azosetshenziswa kuphrojekthi.

Umcwaningi wezokuphepha u-Bar Lanyado, umbhali wocwaningo futhi manje ose-Lasso Security, uthole ukuthi abahlaseli bangakwazi ukulahla kalula iphakethe langempela elinonya endaweni lapho i-ChatGPT ikhomba khona futhi bayinike igama elifanayo nephakheji elikhohliwe. Noma yimuphi unjiniyela olanda iphakheji ngokusekelwe esincomweni se-ChatGPT angase agcine esethule uhlelo olungayilungele ikhompuyutha endaweni yakhe yokuthuthuka.

I-Lanyado's ucwaningo lokulandelela ihlole ukusabalala kwenkinga yephakheji yokuqamba amanga kuwo wonke amamodeli ezilimi ezine ezinkulu: GPT-3.5-Turbo, GPT-4, Gemini Pro (owayekade eyi-Bard), kanye ne-Coral (Cohere). Uphinde wahlola ukukwazi kwemodeli ngayinye ukukhiqiza amaphakheji ahlotshaniswa nemibono ngezilimi ezihlukene zokuhlela kanye nemvamisa abakhiqiza ngayo iphakheji efanayo ekhohliwe.

Ezivivinyweni, i-Lanyado ihlanganise uhlu lwezinkulungwane zemibuzo ethi "ukwenziwa kanjani" onjiniyela abasezindaweni ezihlukene zokuhlela - i-python, node.js, go, .net, ruby ​​- evamise ukufuna usizo kuma-LLM ezindaweni ezithuthukayo. U-Lanyado wabe esebuza imodeli ngayinye umbuzo ohlobene nekhodi kanye nesincomo sephakheji elihlobene nombuzo. Ubuye wacela imodeli ngayinye ukuthi income amanye amaphakheji ayi-10 ukuxazulula inkinga efanayo.

Imiphumela Ephindaphindayo

Imiphumela yayikhathaza. I-64.5% emangazayo "yezingxoxo" uLanyado abe nayo ne-Gemini ekhiqize amaphakheji ahlotshaniswa ne-allucified. Nge-Coral, lelo nani lalingama-29.1%; amanye ama-LLM afana ne-GPT-4 (24.2%) kanye ne-GPT3.5 (22.5%) awazange abe ngcono kakhulu.

Lapho u-Lanyado ebuza imodeli ngayinye isethi efanayo yemibuzo izikhathi eziyi-100 ukuze abone ukuthi amamodeli azoqapha kangaki amaphakheji afanayo, wathola ukuthi amazinga okuphindaphinda nawo aphakamisa amashiya. I-Cohere, isibonelo, yakhafula amaphakheji afanayo ahlotshaniswa ne-hallucated esikhathini esingama-24%; Xoxa nge-GPT-3.5 kanye ne-Gemini cishe ku-14%, ne-GPT-4 ku-20%. Ezimweni eziningana, amamodeli ahlukene abonisa amaphakheji afanayo noma afanayo. Inani eliphakeme kakhulu lamamodeli ahlotshiswe ngehholo eliphambanayo lenzeke phakathi kwe-GPT-3.5 ne-Gemini.

U-Lanyado uthi noma ngabe abathuthukisi abahlukene babuze i-LLM umbuzo ngesihloko esifanayo kodwa baklame imibuzo ngendlela ehlukile, kunethuba lokuthi i-LLM ingancoma iphakheji efanayo ekhohliwe esimweni ngasinye. Ngamanye amazwi, noma yimuphi umthuthukisi osebenzisa i-LLM ukuze athole usizo lokufaka amakhodi angahlangabezana namaphakheji amaningi afanayo.

"Umbuzo ungahle uhluke ngokuphelele kodwa ngesihloko esifanayo, futhi ukubona izinto ezingekho kuye kwenzeka, okwenza le ndlela iphumelele kakhulu," kusho uLanyado. "Ocwaningweni lwamanje, sithole 'amaphakheji aphindaphindayo' emibuzo nezihloko eziningi ezahlukahlukene ngisho nakuzo zonke izinhlobo ezahlukene, okwandisa amathuba okuthi la maphakheji ahlotshaniswe nemibono angabonakali ukuthi asetshenziswe."

Kulula Ukuxhaphaza

Umhlaseli ohlome ngamagama amaphakeji ambalwa akhonjwe inkohliso, isibonelo, angalayisha amaphakheji anamagama afanayo kumakhosombe afanelekile azi ukuthi kunethuba elihle lokuthi i-LLM ingakhomba onjiniyela kuyo. Ukukhombisa ukuthi usongo aluyona inkolelo-mbono, u-Lanyado wathatha iphakethe elilodwa elahlotshaniswa nelokukhohlisa elibizwa ngokuthi “huggingface-cli” ahlangabezane nalo ngesikhathi sokuhlolwa kwakhe futhi walayisha iphakheji elingenalutho elinegama elifanayo endaweni yokugcina ye-Hugging Face yamamodeli okufunda omshini. Onjiniyela balande lelo phakheji izikhathi ezingaphezu kuka-32,000, uthi.

Ngokombono womlingisi osongelayo, ukuqagela kwephakheji kunikeza i-vector eqondile yokusabalalisa uhlelo olungayilungele ikhompuyutha. “Njengoba [sibonile] emiphumeleni yocwaningo, akunzima kangako,” usho kanje. Ngokwesilinganiso, wonke amamodeli ahlonishwe ndawonye ngama-35% emibuzweni ecishe ibe ngu-48,000, kwengeza u-Lanyado. I-GPT-3.5 yayinephesenti eliphansi kakhulu lokubuka izinto ezingekho; UGemini uthole amaphuzu aphezulu, ngokuphindaphinda okumaphakathi okungu-18% kuwo wonke amamodeli amane, uyaphawula.

I-Lanyado iphakamisa ukuthi onjiniyela baqaphe lapho benza ngokuvumelana nezincomo zephakheji ezivela ku-LLM lapho bengenaso isiqiniseko esiphelele sokunemba kwayo. Uphinde athi lapho onjiniyela behlangabezana nephakheji yomthombo ovulekile abangayijwayele kudingeka bavakashele inqolobane yephakeji futhi bahlole usayizi womphakathi wayo, amarekhodi alo okulungisa, ubungozi bawo obaziwayo, kanye nezinga lakhona lokuzibandakanya. Onjiniyela kufanele futhi baskene iphakheji ngokucophelela ngaphambi kokuyethula endaweni yokuthuthukiswa.

• I-SEO Powered Content & PR Distribution. Khuliswa Namuhla.
• I-PlatoData.Network Vertical Generative Ai. Zinike Amandla. Finyelela Lapha.
• I-PlatoAiStream. I-Web3 Intelligence. Ulwazi Lukhulisiwe. Finyelela Lapha.
• I-PlatoESG. Ikhabhoni, I-CleanTech, Amandla, Environment, Ilanga, Ukuphathwa Kwemfucuza. Finyelela Lapha.
• I-PlatoHealth. I-Biotech kanye ne-Clinical Trials Intelligence. Finyelela Lapha.
• Source: https://www.darkreading.com/application-security/pervasive-llm-hallucinations-expand-code-developer-attack-surface