# Amafayela emodeli yeHugging Face angasebenzisa ikhodi efihliwe

Published: 2026-10-12

Amaphakheji emodeli anonya angasebenzisa ikhodi ngesikhathi sokulayishwa okungaphephile. Izibonelo ezintathu ezibhaliwe zichaza ama-Pickle reverse shells, ukuziphatha kweskena sefayela eliphukile kanye nokudideka kwegama lemodeli okukhonjiswe ucwaningo, ngemikhawulo ecacile ezimangweni zokweba imininingwane.

Canonical: https://thedailydiff.dev/zu/video/2026-10-12-huggingface-model-malware/

## Lokho okuhlanganiswa yile vidiyo

- Isisindo semodeli esingamanani, ukulungiswa okusebenzisekayo kanye nekhodi ye-repository yangokwezifiso yizinto ezahlukene.
- I-JFrog ne-ReversingLabs babhale phansi izinkokhelo ze-reverse-shell; imibiko yabo ekhethiwe ayikuqinisekisi ukwebiwa kwemininingwane kubantu abaqanjwe amagama.
- I-Checkmarx Model Confusion kwakuyidemonstration yocwaningo edinga imodeli ekude ehambisanayo kanye nokulayishwa okungaphephile.
- Khetha ama-safetensors nezilayishiwe ezilungisiwe, buyekeza futhi ubeke ikhodi yangokwezifiso, futhi uhlukanise amaphakheji angajwayelekile ezimfihlo zomuntu siqu kanye nokufinyelela kwenethiwekhi okungadingekile.

## Izahluko

- 0:00 Ulungiselelo lwemodeli lungamema kanjani isivakashi?
- 0:35 Ulayishi kanjani isilayishi imiyalelo yephakethe?
- 1:09 Yiziphi izimfihlo ezifinyeleleka inqubo?
- 1:38 Kungani indlela yendawo elahlekile ingaba imodeli ekude?
- 1:55 Amaphakheji aphukile akugwema kanjani iskena esidala?
- 2:44 Yini egcina ukhiye ngaphandle kokufinyelela kwesivakashi?
- 3:12 Kungani ngingawubeka lolu hlolo olusha?

## Umbhalo obhaliwe ohunyushwe

Kuhunyushwe kusuka ekulandiseni kokuqala kwesiNgisi. Umsindo namazwibela atholakalayo alawulwa yi-YouTube.

### Ulungiselelo lwemodeli lungamema kanjani isivakashi?

0:00 Ifayela lemodeli liyinombolo nje kuphela ku-disk. Ukulayisha iphakheji elilodwa elinonya leHugging Face kwavula i-remote shell, okuyincazelo enesifiso ye-AI yendawo. Ukulayisha kuba ukusebenza kanjani? Yiziphi izimfihlo ezingazifinyelela? Yini empeleni ekumisa? Futhi gcina iphazili elilodwa engqondweni. Imodeli ingahluleka ukulayisha ngemuva kokuba ingxenye eyingozi isivele yenzekile.

0:18 Lena yiThe Daily Diff, ngaphansi kwezinto. I-Hugging Face iyisikhulumi sokwabelana ngamamodeli wokufunda ngomshini, amasethi edatha nezinhlelo zokusebenza. Isisindo semodeli yizinombolo ezifundiwe. I-repository ingaqukatha futhi ukucushwa nekhodi ye-Python. Ingozi lapha ukupakisha okusebenzisekayo okuwela kukhompyutha yakho lapho uyi ulayisha. I-Pickle igcina izinto ze-Python ngemiyalelo yokuzakha kabusha.

### Ulayishi kanjani isilayishi imiyalelo yephakethe?

0:37 Ulayishi ongaphephile angalandela leyo miyalelo bese ebiza imisebenzi ngenkathi evula. Ukulanda ama-bytes kubeka iphakethe ku-disk yakho. Ukuyivula kuyivula. Ngenye indawo, uhlobo lwefayela luthuthukiselwe kumphathi wesistimu ongafundiwe. I-JFrog yabika i-Goober Two ngoFebhuwari 2024. Inkokhelo yayo ye-Pickle yavula i-reverse shell, exhuma ngaphandle ukuze inikeze ukufinyelela okude komyalo. I-repository yasuswa. I-JFrog iphinde yathola izinkomba ezisikisela ukubandakanyeka kwabacwaningi,

1:02 ngakho-ke ukubhaliswa akucacile. Isivakashi esingafuneki sabhalwa. Ukwebiwa kwemininingwane kubantu abayizisulu akuzange kuqinisekiswe.

### Yiziphi izimfihlo ezifinyeleleka inqubo?

1:09 Ikhodi esebenza njengomsebenzisi wakho ingafunda noma yini leyo nqubo engayifunda. Ukhiye wefu ofakiwe noma ithokheni elifinyelelekayo liba inhloso engaba khona. Ukuyithumela ngaphandle nayo idinga indlela yenethiwekhi esebenzayo. Isisindo samanani kuphela asinawo amandla okhiye bemfihlo. Izimvume zakho zinikeza ingxenye ethakazelisayo yekhadi lemodeli. Kunomunye umnyango. Amanye amamodeli akucela ukuthi uthembe ikhodi ekude, egunyaza i-Python ku- repository. Amafayela e-safetensor angahlala eduze kwalowo khodi.

1:32 Ukuhlola ifomethi yesisindo kuhlanganisa indawo eyodwa, kuyilapho ikhodi ye-repository isasilinde ukuthi uyinike isivakashi isivakashi isivakashi isivakashi.

### Kungani indlela yendawo elahlekile ingaba imodeli ekude?

1:38 I-Checkmarx yabeka i-Model Confusion. Indlela yendawo yemodeli elahlekile ingaxazululela i-repository yomphakathi enegama elifanayo. Hlanganisa lokho nekhodi ekude enikwe amandla, futhi iphakheji elingalungile lingasebenza. Lena kwakuyidemonstration yocwaningo ebandakanya isampula lenkampani engaqanjwanga igama, elilungiswe ngaphakathi kosuku. Ifolda elahlekile yathola umnikazi wendawo. I-ReversingLabs yabika amasampula amabili ngoFebhuwari 2025,

### Amaphakheji aphukile akugwema kanjani iskena esidala?

1:58 ebiza inqubo ngokuthi i-nullify AI. Zombili ziqukethe izinkokhelo ze-reverse shell. Ukupakisha kwazo kwe-seven zip kuvimbe ukulayisha okujwayelekile kwe-PyTorch, ngakho le ndlela yadinga ukukhishwa kanye nokulungiswa okungaphephile. Ama-Pickle streams aphukile aphinde adalula indawo engabonakali yeskena. Iskena esihlulekile sidinga ukuphenywa ngaphambi kokuba noma ubani avule iphakethe. Ekuhlolweni okungenabungozi kwabacwaningi, imiyalelo yabhala ifayela ngaphambi kokuba isilayishi

2:18 sifinyelele engxenyeni ephukile futhi sakhipha iphutha. Cabanga uvula iphakethe, ukhulule inhlwazi, bese uqaphela ukuthi ilebula lokuthumela alingavumelekile. Umyalezo wephutha ufika ngegrama elihle kakhulu kanye nesikhathi esibi kakhulu. I-Hugging Face yasusa amasampula abikiwe phakathi kwamahora angama-24 futhi yashintsha iskena. Lelo thola elithile lenza ukulungiswa. Iskena sokuphepha siwubufakazi obuwusizo, futhi ukuphatha wonke amaphutha eskena njenge uphawu lokuphepha kuhlala kuyincazelo enesandla esivulekile kakhulu yokwehluleka.

2:43 Khetha ama-safe tensors nezilayishiwe ezilungisiwe zamanje.

### Yini egcina ukhiye ngaphandle kokufinyelela kwesivakashi?

2:46 I-PyTorch yaqhubeka nesilayishi sayo esizenzakalelayo kusukela enguqulweni 2.6. Buyekeza ikhodi yangokwezifiso bese ubeka isibuyekezo esibuyekeziwe. Qalisa amaphakheji angajwayelekile ngokuhlukana ngaphandle kokhiye bomuntu siqu, ngokufinyelela okulinganiselwe kwenethiwekhi. Isiqukathi esigcwele izimvume ezifakiwe siyibhasikidi lesipho elinamaspace. Le phazili yokuqala inempendulo. Ukuhluleka kokulayisha kukutshela ukuthi ukusebenza kumile kuphi. Imiyalelo yangaphambili ingase isebenzile kakade.

3:08 Gcina izimfihlo ngaphandle kokufinyelela kwayo ngaphambi kokuvula iphakethe. Isigwebo, ngaphansi kwezinto.

### Kungani ngingawubeka lolu hlolo olusha?

3:16 isilayishi. Ukulanda kwamahhala kufanele kuze nemodeli, futhi uhlu lwezivakashi kufanele kuhlale kuyisinqumo sami. Uma uthanda ukufunda lokhu kunokuba ungizwe ngisho, i-diff ifika ebhokisini lakho leposi njalo ekuseni, mahhala ku-the daily diff dot dev, isixhumanisi esingezansi. Futhi yilokho okuhlukile kwanamuhla. NginguNiko ovela kwa-Axrisi. Hlanganisa ngokuzibophezela.

## Imithombo

- [Data Scientists Targeted by Malicious Hugging Face ML Models with Silent Backdoor](https://jfrog.com/blog/data-scientists-targeted-by-malicious-hugging-face-ml-models-with-silent-backdoor/) — JFrog
- [Malicious ML models discovered on Hugging Face platform](https://www.reversinglabs.com/blog/rl-identifies-malware-ml-model-hosted-on-hugging-face) — ReversingLabs
- [AI Model Confusion: An LLM/AI Model Supply Chain Attack](https://checkmarx.com/zero-post/hugs-from-strangers-ai-model-confusion-supply-chain-attack/) — Checkmarx Zero
- [Pickle Scanning](https://huggingface.co/docs/hub/en/security-pickle) — Hugging Face
- [Safetensors](https://huggingface.co/docs/safetensors/en/index) — Hugging Face
- [Loading custom models](https://huggingface.co/docs/transformers/en/models) — Hugging Face
- [Serialization semantics](https://docs.pytorch.org/docs/2.14/notes/serialization.html) — PyTorch