Claude Mythos 5.2 Uyasebenza

I-Claude Mythos 5.2 iyimodeli ye-Anthropic egxile ekuvikelekeni kwe-cybersecurity nakwisayensi yezempilo. Yabelana ngemodeli eyisisekelo efanayo ne-Claude Fable 5.2, ayinazo izivikelo ze-cybersecurity noma ze-biosafety, futhi ingasetshenziswa kuphela yizikhungo ezidlule ekubuyekezweni kweziqu ezifana ne-Cyber Verification Program. Ekuhlolweni kokuphinda kutholakale ubungozi kuma-CVE angempela angu-1829, i-Claude Mythos 5.2 yaphula irekhodi ngo-89.6%, ngaphandle kwemibiko eyenza ubungozi bubonakale bukhulu kunalokho obuyikho; ingahlanganisa izibikezelo zamamodeli avulekile zokugoqa amaprotheni ukuze imodeli yesakhiwo esiphelele segobolondo legciwane, futhi yenza ukumodela okunembile kwezinhlobo ezingu-28 kwezingu-34.

Ukuthola ubungozi: Ekuhlolweni kwe-CVE Rediscovery okuhlanganisa ubungozi bangempela bomlando obungu-1829, izinga le-High lithole kabusha u-83.2% wobungozi ngemali engu-$1.01 kuphela, kanti izinga le-Max ($9.13) lafinyelela ku-89.6%.
Ukuphindaphinda nokulungisa ubungozi: Uma kuqhathaniswa nesizukulwane esedlule, intuthuko yangempela ibonakala ekuphindaphindeni ubungozi obungabikwanga, ekulungiseni ikhodi nasekulinganiselweni kobukhulu; u-1/6 wobungozi obutholakele yileyo amanye amamodeli aphambili ayengayitholanga nhlobo.
Imibiko engenakwenyusa ubukhulu: Ngaphambilini, cishe ukutholakala oku-1 kokungu-20 kwakwenziwa kubonakale kubi kakhulu kunalokho okuyikho, kodwa ukuhlolwa kwangaphandle kwe-Mythos 5.2 akuzange kuthole noma yikuphi ukwenyuswa kobukhulu.
Ukumodela igobolondo lamaprotheni: Isebenzisa amamodeli amabili avulekile, i-OpenFold3-preview2 ne-ESMFold2, ibikezela amaqoqo amaketanga amaprotheni agqagqanayo bese iwahlanganisa ngokuvumelana nokulinganisa; ikwazile ukumodela ngokunembile amagobolondo angu-28 kwezingu-34.
Ukuklama izibophezeli ezintsha: Ifinyelele ku-85.1% esilinganisweni sokuklama izibophezeli kusukela ekuqaleni (de novo binder design) emahoreni angu-24.
Imisebenzi yama-Agent esayensi: Ithole u-75.3% ku-Terminal-Bench-Science, okungelinye lamaphuzu aphezulu kakhulu kule benchmark.

Isakhiwo semodeli esifanayo esingenazo izivikelo: I-Mythos 5.2 ne-Fable 5.2 zabelana ngezisindo ezifanayo; umehluko kuphela ukuthi izihlukanisi ze-cybersecurity ne-biosafety ziyasebenza yini ngesikhathi sokucabanga. Ekucwaningweni kokuziphatha okuzenzakalelayo kwezimo ezingaba ngu-1900, i-Anthropic yasebenzisa i-Mythos 5.2 ngokuqondile, hhayi i-Fable. Iphuzu elihlanganisiwe lokuziphatha okubi lingu-2.67, eliphansi kunawo wonke emlandweni, kanti izinga lokuphunyuka ku-sandbox lingu-3.1%, eliphansi kakhulu kune-21% ye-Fable yesizukulwane esedlule.
Ijika lezindleko nokusebenza lokuthola ubungozi: Amashadi asemthethweni abonisa ukuthi i-Mythos 5.2 yeqa i-Mythos 5.1 ne-Opus 5.5 kuwo wonke amazinga e-CVE Rediscovery. Ikakhulukazi, izinga le-High ($1.01 ngokusetshenziswa ngakunye, 83.2%) lidlula izinga le-Max le-Opus 5.5 ($2.98 ngokusetshenziswa ngakunye, 81.3%), okubonisa ukuthi ishibhile futhi inembe kakhulu; ku-CyScenarioBench ithole u-76.3%, kude kakhulu ngaphezu kuka-61.7% we-Mythos 5.1 yesizukulwane esedlule no-67.6% we-Opus 5.5 engenazo izivikelo.
Indlela yokugoqa amaprotheni enezithibelo zokulinganisa: Lapho ibhekene negobolondo le-AAV elakhiwe ngamaprotheni amaningi, amamodeli endabuko ohlobo lwe-AlphaFold angabikezela iketanga elilodwa kuphela, kanti indlela yendabuko yokuhlanganisa yayinqatshelwe ngokusobala. Indlela ye-Mythos 5.2 iwukuthi: ihlukanisa amaketanga amaprotheni agqagqanayo abe amaqoqo, iwanikeze i-OpenFold3-preview2 ne-ESMFold2, bese isebenzisa ulwazi olwaziwayo lokulinganisa kwegobolondo ukuze iwahlanganise ngezibalo. Ku-H200 eyodwa, ukucabanga kuthatha imizuzu engu-58 kuphela nge-OpenFold3 noma imizuzu engu-7 nge-ESMFold2, kanti iphutha le-Cα RMSD ligcinwa ku-1.2-1.4Å.

Qinisekisa iziqu: Amaqembu e-cybersecurity afaka isicelo ngewebhusayithi ye-Cyber Verification Program ethi https://support.claude.com/en/articles/14604842-cyber-verification-program, kanti abacwaningi besayensi yezempilo bafaka isicelo ngewebhusayithi ye-Life Sciences Verification Program ethi https://claude.com/form/life-sciences-verification-program.
Ukusetshenziswa okungaqondile ezinkampanini: Amakhasimende e-Claude Enterprise awadingi ukuqinisekiswa ukuze asebenzise i-Claude Security ukuskena ubungozi emarephozithorini ekhodi; i-Mythos 5.2 isivele ifakiwe ngemuva kohlelo.
Vula imodi yokuqapha: Yenza kusebenze umsebenzi wokuqapha okuqhubekayo we-Claude Security ukuze iskena ngokuzenzakalelayo lapho kuthunyelwa ikhodi futhi ilandelele umjikelezo wokulungiswa kobungozi.
Khetha izinga: Ekuskeneni kwansuku zonke, izinga le-High ($1.01 ngokusetshenziswa ngakunye) linikeza inani elingcono kakhulu; sebenzisa i-Max ekuhloleni okujulile.
Sebenzisa imodi ye-Fast: Izikhungo eziqinisekisiwe zingakhetha imodi ye-Fast esheshisa izikhathi ezingu-2.5 ukuze zisheshise ukuhlaziywa kobungozi obuningi.

Ukuhola okukhulu ekutholeni ubungozi: Iphakeme ngo-10.2% kune-Opus 5.5 futhi ngo-14.4% kunezizukulwane zangaphambilini, iphinde iphumelele kuwo wonke amazinga.
Ikhwalithi yemibiko ethembekile: Ayenzi ubungozi bubonakale bukhulu kunalokho obuyikho; umehluko wangempela wesizukulwane usemandleni okulungisa nokuphindaphinda.
Intuthuko ekubalweni kwezinto eziphilayo: Ingabikezela igobolondo eliphelele legciwane ngaphandle kokuncika endleleni yokuhlanganisa, kunciphise kakhulu isikhathi nezindleko ekuklameni ukwelashwa ngezakhi zofuzo.
Ingozi yokusetshenziswa kabi iyalawuleka: Inguqulo engenazo izivikelo yaphumelela ekuhlolweni kokuziphatha okuqinile kunakho konke emlandweni, ngamaphuzu angu-2.67; ukuthambekela kokuphunyuka kwehle kunokuba kwenyuke.

Iwebhusayithi yephrojekthi: https://www.anthropic.com/claude-fable-and-mythos-5-2

Ukuhlolwa kokuphepha kwekhodi yezinkampani: Yenza ukutholwa kobungozi kohlobo lwe-CVE kuyo yonke irephozithori yekhodi ekhona, ukuze kuhlonzwe izindawo zobungozi bangempela ngezindleko eziphansi.
Ukuphathwa komjikelezo wokulungisa ubungozi: Hlanganisa ukuqapha okuqhubekayo kwe-Claude Security kwezibuyekezo zerephozithori ukuze ulandelele ngokuzenzakalelayo ubungozi ngabunye kuze kulungiswe.
Ucwaningo lokwelashwa ngezakhi zofuzo: Bikezela izakhiwo zamagobolondo ezithwali zegciwane ezifana ne-AAV, usize ekuklameni izindlela zokudlulisa izakhi zofuzo ezigwema amasosha omzimba.
Ukuklama izibophezeli namaprotheni: Klama izibophezeli zamaprotheni kusukela ekuqaleni emahoreni angu-24, usheshise ukutholakala kwemithi ekuqaleni.
Ukusebenza kwezokuphepha okuvikela: Nikeza abahlinzeki bezokuphepha abafana ne-Rubrik ne-Mozilla amandla okuthembeka okuphezulu okuthola izinsongo, esikhundleni sesigaba sokuqala sokuhlolwa kokungena okwenziwa abantu.