Imodeli entsha yobuhlakani bokwenziwa ye-Anthropic, eyaziwa ngokuthi iClaude Mythos Preview , isibe yisizinda sempikiswano yomhlaba wonke ngemikhawulo ye-AI ethuthukisiwe. Le nkampani ngokwayo iyavuma ukuthi uhlelo lunamandla kakhulu maqondana nokuphepha kwe-cyber kangangokuthi inqume ukungalukhiphi kabanzi, isinqumo esingavamile emkhakheni ojwayele ukuziqhayisa ngentuthuko ngayinye entsha.
Okusengozini akukhona nje ukuthuthuka okuqhubekayo kunamamodeli angaphambilini, kodwa ukwenyuka okubonakalayo ekhonweni lokubona nokusebenzisa ubuthakathaka bekhompyutha . Ohulumeni, amabhange amakhulu, izikhungo zezimali ezinkulu, kanye nabalawuli baseYurophu baqapha isimo eduze, beqaphela ukuthi ithuluzi elinjalo lingaqinisa ukuzivikela kwezinhlelo ezibalulekile , kodwa futhi livule umnyango wokuhlaselwa okukhulu kakhulu uma liwela ezandleni ezingafanele.
Iyini ngempela iClaude Mythos futhi kungani ukwethulwa kwayo kuye kwabambezeleka?
I-Claude Mythos ingenye yamamodeli amasha emndenini wakwaClaude, i-Anthropic's AI ecosystem encintisana ne -OpenAI's ChatGPT kanye ne-Google's Gemini . Iyimodeli yenhloso ejwayelekile ekwazi ukucabanga, ukuhlela, kanye nokusebenza ngomongo wesikhathi eside, kodwa isici sayo esiphikisana kakhulu ukusebenza kwayo ekuvikelekeni kwe-cyber okuhlaselayo nokuzivikela.
Lokho okubizwa ngokuthi “amaqembu abomvu ,” ochwepheshe abahlola izinhlelo ze-AI ngemikhawulo yabo, baphetha embikweni wangaphakathi ngokuthi i-Mythos “inekhono elimangalisayo” emisebenzini yokuphepha kwe-cyber. Ezivivinyweni zokulinganisa ezifana ne -SWE-bench Verified kanye ne-SWE-bench Pro , ezenzelwe ukukala ikhono lokuxazulula izinkinga zobunjiniyela besofthiwe yangempela, imodeli yenza kahle kakhulu kunezinye izindlela zezentengiselwano ezihamba phambili, kufaka phakathi izinguqulo ezithuthukisiwe ze-GPT kanye ne-Gemini, ngokusho kwedatha enikezwe yi-Anthropic uqobo.
Ngale kwezilinganiso, okuye kwaphakamisa izinkomba zokuxwayisa ukuthi i-Mythos ikwazile ukuthola ubuthakathaka bezinsuku ezingenalutho —amaphutha ayengaziwa ngaphambili—kusofthiwe esetshenziswa kabanzi, eminye yayo ineminyaka engaphezu kwamashumi amabili ubudala. Ezinhlelweni ezifana nezingxenye ze-OpenBSD, FFmpeg, kanye ne-FreeBSD, imodeli ayizange igcine ngokuthola amaphutha ayengabonwa iminyaka kodwa futhi yakhiqiza nokuxhashazwa kokusebenza ukuze iwasebenzise.
Ebhekene nale miphumela, i-Anthropic yakhetha isinqumo esingavamile embonini: ukwethula imodeli bese imemezela ngokushesha ukuthi ngeke imakethwe obala, icabanga ukuthi ibeka izingozi zokuphepha kwe-inthanethi ezingakaze zibonwe. Le nkampani iphikelela ngokuthi i-Mythos iyimodeli "ehambisana kahle kakhulu" abayikhile, kodwa iyavuma ukuthi amandla ayo amakhulu andisa imiphumela yanoma yikuphi ukusetshenziswa kabi.
Imodeli enamakhono "okuba yi-hacker" adlula kakhulu amakhono abantu
Imibhalo yobuchwepheshe kanye nemibiko evela ezinhlanganweni ezahlukene iyavuma ukuthi i-Mythos iphawula inguquko ezenzakalelayo zokuhlaselwa okuyinkimbinkimbi . Ezindaweni zokuhlola ezilingisa amanethiwekhi ezinkampani zangempela, uhlelo lwalukwazi ukuhlanganisa ubuthakathaka, lwandise amalungelo, futhi lwafinyelela ukufinyelela okuqhubekayo ngamahora—imisebenzi eyayizothatha izinsuku noma amasonto ochwepheshe bomuntu.
Isibonelo, enjinini yeJavaScript yeFirefox, izinguqulo zangaphambilini zamamodeli e-Anthropic azikwazanga ukuguqula amaphutha abe yizenzo zokusebenza. I-Mythos, ngaphansi kwezimo ezifanayo zokuhlola, ikhiqize inqwaba yezenzo zokusebenza , iphinda ngokunembile ukusetshenziswa kwamavektha asebenza kahle kakhulu. Kumapulatifomu okuhlaziya afana ne-OSS-Fuzz, aklanyelwe ukuthola amaphutha kusofthiwe yomthombo ovulekile, idunyiswa ngokuthola ubuthakathaka obukhulu obungazange buqashelwe naphezu kweminyaka yokuhlolwa okuzenzakalelayo.
Imodeli iphinde yabonisa amakhono amangalisayo obunjiniyela obuphambene : ingakha kabusha ingxenye ye-logic yohlelo kusuka kuma-binary ahlanganisiwe, bese ithola futhi isebenzise ubuthakathaka ngaphandle kokufinyelela ikhodi yomthombo wokuqala. Lolu hlobo lwekhono lusondeza i-AI ezimweni, kuze kube muva nje, ezazibhekwa njengesizinda esikhethekile samaqembu abantu akhethekile kakhulu.
Esinye seziqephu ezivame ukucashunwa ekuhlolweni kokuphepha yilokho okubizwa ngokuthi "ukuhlolwa kwesandwich." Endaweni yelabhorethri ehlukanisiwe, iMythos yanikezwa ukulawula uhlelo olunemiyalelo ecacile yokuzama ukuphunyuka ebhokisini lesanti futhi ixhumane nomcwaningi ophethe ukuhlolwa. Imodeli yasebenzisa ngempumelelo uchungechunge lobuthakathaka ukuze ishiye indawo yayo evinjelwe futhi ithumele i-imeyili kumphathi, owayengekho ehhovisi ngaleso sikhathi. Nakuba lesi sigameko senzeke enguqulweni yangaphakathi yangaphambilini futhi ngaphansi kwemiyalo eqondisiwe, sibonisa izinga uhlelo olungasebenza ngalo ezimweni eziyinkimbinkimbi ngaphandle kokubhekwa okuncane.
Naphezu kwalezi ziboniso, abahlaziyi baphikelela ekucaciseni ukuthi asibhekene ne-AI "eqondayo" noma umuntu onenkululeko yokuzikhethela . I-Mythos ayizinqumeli ngokwayo ukuhlasela izinhlelo: yenza imisebenzi eceliwe ukuba iyenze kahle ngangokunokwenzeka. Ingozi, ngalomqondo, akukhona ukuthi imodeli izovukela, kodwa ukuthi othile uzoyisebenzisa—noma ayiphoqe ngokusebenzisa izikhuthazo eziyinkimbinkimbi—ukwenza izenzo ezilimazayo.
Iphrojekthi ye-Glasswing: Izinganekwane zisebenza njengokuzivikela… kwabambalwa abakhethiwe
Esikhundleni sokwenza i-Mythos itholakale emphakathini, i-Anthropic ikhethe ukuyizungeza ngohlelo oluthile, i-Project Glasswing , oluklanyelwe ukusebenzisa amakhono alo modeli ngendlela elawulwayo yokuvikela isofthiwe ebalulekile. Lesi sinyathelo sihilela ukunikeza uhlelo, ngaphansi kwezimo eziqinile zokusetshenziswa, eqenjini elikhethiwe lezinkampani ezinkulu zobuchwepheshe, abahlinzeki bengqalasizinda, kanye nezikhungo zezezimali.
Phakathi kwezinhlangano ezinokufinyelela kukhona iziqhwaga ezifana ne -Amazon Web Services, i-Apple, i-Microsoft, i-Google Cloud , i-Nvidia, kanye ne-Broadcom , kanye nezinkampani zokuphepha kwe-inthanethi ezifana ne-CrowdStrike, isofthiwe yazo enephutha eyabangela ukuphazamiseka okukhulu emhlabeni wonke ngo-2024. Lezi zijoyinwa amabhange adumile emhlabeni wonke, okuhlanganisa i-JP Morgan Chase kanye namaqembu amaningana amakhulu e-Wall Street , ngaphezu kwezinye izinhlangano ezinesibopho sokugcina ingqalasizinda ye-IT ebucayi.
I-Anthropic iphinde yamemezela uxhaso lwemali oluyizigidi eziyi-100 zamaRandi ukuze lezi zinhlangano zisebenzise i-Mythos ekuhlaziyeni ubungozi, kanye neminikelo ezisekelweni ezivulekile njenge-Linux Foundation kanye ne-Apache Software Foundation. Umgomo osemthethweni ucacile: ukuvumela labo abaphatha isofthiwe ebaluleke kakhulu emhlabeni ukuthi babone futhi balungise ubungozi ngaphambi kokuba amathuluzi anjalo atholakale kubahlaseli abangaba khona.
Nokho, leli qhinga libangela ukungakhululeki okuthile ngaphakathi komkhakha. Ngakolunye uhlangothi, liqinisa umqondo wokuthi ubuchwepheshe buyingozi ngokwanele ukuthi budinga ukufinyelela okulinganiselwe. Ngakolunye uhlangothi, kudala ukwahlukana phakathi kwalabo abazuza "esivikelweni" se-Mythos nalabo abashiywe ngaphandle . Izinkampani nezinhlangano zikahulumeni ezingeyona ingxenye ye-Glasswing zisengozini yokubhekana nobuthakathaka obutholakale futhi balungiswa ezindaweni ezinelungelo kodwa zihlala zikhona ezinhlelweni zazo.
EYurophu, lokhu kungalingani kukhathaza kakhulu labo abaphethe ingqalasizinda ebalulekile kanye namaqembu okuphepha amaqembu amakhulu ezimboni kanye nezezimali, ababheke eduze ukuthi iBrussels kanye nenhloko-dolobha yaseYurophu zingaqinisekisa yini ukuthi izinhlelo ezifanayo zifaka abadlali ababalulekile kuleli zwekazi ngokulinganayo kanye nobukhosi bamafu nabalingani babo base-US.
Impendulo evela kohulumeni, abalawuli kanye nomkhakha wezezimali
Umthelela we-Mythos udlulela ngale komkhakha wezobuchwepheshe. Ezinsukwini ezimbalwa, isimemezelo sale modeli saqala imihlangano ephezulu e-United States naseYurophu . UNobhala Wezezimali wase-US wabiza izinhloko zamabhange amakhulu ezwe eWashington ukuze ahlole izingozi ezingaba khona uhlelo olungazifaka ekuzinzeni kwezezimali, kanti uSihlalo we-Federal Reserve naye wahlanganyela kulezi zingxoxo.
Ngokusho kokuvuza okubikwe ngabezindaba bamazwe ngamazwe, lezi zinhlangano zakhuthazwa ukuthi zihlole i-Mythos ngendlela yokuzivikela , ziyisebenzisele ukuskena ingqalasizinda yazo ukuze zithole ubuthakathaka ngaphambi kokuba abanye bakwazi. Umyalezo oyimfihlo uwukuthi usongo lubi ngokwanele ukuthi lufanele impendulo ehlangene phakathi kwemikhakha yomphakathi neyangasese.
Okwamanje, umsunguli we-Anthropic uqinisekisile ukuthi inkampani isezingxoxweni eziqondile nohulumeni wase-US mayelana ne-Mythos kanye namamodeli esikhathi esizayo. Lezi zingxoxo zenzeka ngaphansi kwesimo esishubile, kulandela ukufakwa kwamuva kwenkampani ohlwini lwezingozi zochungechunge lokuhlinzekwa yiziphathimandla zase-US , ngemuva kokuvela kwengxabano mayelana nokusebenzisa koMnyango Wezokuvikela amamodeli awo.
Ngaphesheya kwe-Atlantic, i-European Union iqaphele. IKhomishini yaseYurophu ivumile obala indlela eqhubekayo neqaphile yamamodeli afana ne-Mythos, kanti abalawuli bezezimali e-UK nakuwo wonke amazwekazi sebeqalile ukutadisha ngqo imiphumela yayo engaba khona emabhange nasezimakethe. I-AI Security Institute (AISI) kahulumeni wase-UK ichaze lolu hlelo njengesinyathelo esibalulekile maqondana nokuphepha kwe-cyber uma kuqhathaniswa nezizukulwane ezedlule.
ESpain, nakuba impikiswano yomphakathi isalinganiselwe, izinhlangano zokuphatha kanye namaqembu okuphepha kwe-inthanethi avela emabhange nasezinkampanini ezinkulu zamandla ziqapha ngokucophelela lezi zinguquko. Emkhakheni wezezimali waseYurophu, noma yikuphi ukuthuthuka okungase kube lula ukuhlaselwa okuhlelekile kwezinhlelo zokukhokha, amanethiwekhi okuxhumana phakathi kwamabhange, noma amapulatifomu okuhweba kuyimbangela yokukhathazeka okukhulu.
Ukungabaza, ukungabaza kanye nempikiswano mayelana "nomqondo ongcolile" ozungeze izinganekwane
Indaba ka-Anthropic, ehlanganisa izixwayiso zokuphepha nezibalo zokusebenza ezinhle kakhulu, ayizange ibe ngaphandle kwabagxeki bayo. Ochwepheshe abaningana be-AI kanye nokuphepha kwe-cyber banxuse ukuthi kuqashelwe ekuchazeni izimangalo zenkampani , bekhomba ukuthi idatha eningi etholakalayo ivela emibikweni yangaphakathi kuphela.
Abanye abahlaziyi babukeze ngokuningiliziwe imibhalo ebanzi eshicilelwe yi-Anthropic futhi bakhomba ukuthi inani "lezinkulungwane zobuthakathaka obukhulu" lisekelwe ekuthathweni kwemininingwane evela enanini elincane lamacala abuyekezwe ngesandla. Kwezinye izingxoxo zokuhlola, kubikwa ukuthi i-Mythos ithole inani elikhulu lamaphutha abucayi, kodwa kude nesimo esicishe sibe yinhlekelele esiphakanyiswe ezinye izihloko.
Ezinye izifundo ezizimele zizame ukuqhathanisa ukusebenza kweMythos namamodeli amancane, avulekile, okunikeza izingcezu zekhodi ezisengozini kuma-AI ahlukene ukuze kubonakale ukuthi angakwazi yini ukuthola amaphutha afanayo. Imiphumela ikhombisa ukuthi amanye amamodeli avulekile nawo ayakwazi ukubona ubuthakathaka obuyinkimbinkimbi , ephonsela inselelo umbono wokuthi iMythos isebenza ngendlela ehluke ngokuphelele kuzo zonke izimo.
Lezi zinhlobo zezibonelo eziphikisayo aziphikisi amakhono e-Mythos, kodwa ziphakamisa ukuthi ingxenye yendaba "eyingozi kakhulu ukuyishicilela" nayo inomkhawulo wokumaketha . Ukwethula imodeli njengenamandla kakhulu kanye nengozi engaba khona kuqinisa isithombe sobuholi bezobuchwepheshe kanye nomthwalo wemfanelo, into eyigugu kakhulu emakethe encintisana kakhulu.
Umlando wakamuva wale mboni ukhumbuza futhi isibonelo se-GPT-2 ngo-2019, lapho i-OpenAI ekuqaleni inquma ukungashicileli imodeli ephelele, ithi iyingozi kakhulu ngenxa yokuthi inamandla okukhiqiza ulwazi olungesilo iqiniso. Ekugcineni, leyo nguqulo yakhishwa emphakathini ngaphandle kokuba noma yiziphi izinhlekelele ezibikezelwe zenzeke, futhi ochwepheshe abaningi bayibhekisele njengesibonelo sokusabela ngokweqile. Nge-Mythos, umehluko ukuthi ukugxila akusekho embhalweni ngokwawo, kodwa ekuthembekeni kwengqalasizinda yedijithali - indawo ebucayi kakhulu kohulumeni namabhange.
Ibhalansi ebucayi phakathi kokuphepha, ibhizinisi, kanye nokufinyelela ubuchwepheshe
Ngaphandle kwenkulumo-ze yabezindaba, isimo se-Mythos siphakamisa umbuzo oyisisekelo: ubani onquma ukuthi imodeli ye-AI iyingozi kakhulu ukuyikhipha nini, futhi ngaphansi kwaziphi izindlela? Okwamanje, lesi sinqumo sithathwe ngohlangothi olulodwa, esenziwe yi-Anthropic, ekhethe ukugcina uhlelo luhlobo lokuvalelwa olulawulwayo, ilugcinele abalingani abakhethiwe.
Lesi simo asisekelwe kuphela ezinkingeni zokuphepha. Ukusebenzisa imodeli enezici zeMythos kubiza kakhulu ngokwezibalo , futhi inkampani ngokwayo iyavuma ukuthi okwamanje ayinayo ingqalasizinda edingekayo yokuyisebenzela kakhulu izigidi zabasebenzisi. Empeleni, izinyathelo zokuphepha kanye nemikhawulo yobuchwepheshe kuyahambisana, okunikeza i-Anthropic isikhathi sokulungisa imodeli kanye nokusetshenziswa kwayo.
Ngesikhathi esifanayo, inkampani isiqalile ukuhlukanisa ngokucacile phakathi kwemikhiqizo yayo ehlukahlukene. Nakuba i-Mythos isalokhu iyindinganiso yangaphakathi ethuthuke kakhulu , egcinelwe ucwaningo kanye nokusebenzisana ngamasu, amanye amamodeli afana ne-Claude Opus 4.7 ahloselwe ukusetshenziswa kwansuku zonke ngamabhizinisi nochwepheshe. I-Anthropic ivumile esidlangalaleni ukuthi i-Opus 4.7 "ayinamandla kangako" kune-Mythos ngokwejwayelekile futhi, ikakhulukazi, maqondana namakhono ayo e-cyber - into engavamile embonini evame ukwethula imodeli ngayinye entsha njengengcono kakhulu kuzo zonke izici.
Kulo modeli, i-Mythos isebenza njengendawo yokuhlola amakhono esizukulwane esilandelayo , kuyilapho amamodeli atholakala kwezentengiselwano ehlanganisa ingxenye kuphela yalawo makhono, enemikhawulo eyengeziwe eyenzelwe ukunciphisa izingozi. Ezinhlanganweni eziningi zaseYurophu ezinentshisekelo yokusebenzisa i-AI ngaphandle kokuba semgqeni ophambili wokudalulwa, lokhu kuhlukaniswa phakathi kwamamodeli "okuhlola" kanye "nokukhiqiza" kungaba yindlela enengqondo, uma nje kukhona ukucaca okwanele mayelana namakhono angempela esistimu ngayinye.
Ekugcineni, okuvelayo yisimo lapho ukuphepha kwe-inthanethi kungena ngokugcwele enkathini ye-AI yokuhlasela nokuzivikela enkulu . Amathuluzi afana ne-Mythos athembisa ukusheshisa ukuhlonza ubuthakathaka ezinhlelweni ebezisebenza iminyaka eminingi, kodwa futhi aphoqa ukucabanga kabusha ngendlela ubuchwepheshe obusekela umnotho wedijithali obusatshalaliswa futhi bubuswa ngayo. EYurophu naseSpain, inselelo ngeke nje ibe ukuzivikela kumamodeli e-AI anamandla kakhulu, kodwa futhi nokuqinisekisa ukuthi awashiywa ngaphandle kwezindlela eziwavumela ukuthi asetshenziswe ukuqinisa ukuphepha kwawo.
