Anthropic adalengeza Claude Mythos, chitsanzo cha AI chomwe chingathe kupanga zinthu zogwirira ntchito.

Anthropic yalengeza za pulojekiti ya Glasswing, yomwe ipereka mwayi wopeza mtundu woyamba wa Claude Mythos AI model yake kuti izindikire zofooka ndikukweza chitetezo cha mapulogalamu ofunikira. Omwe akutenga nawo mbali mu pulojekitiyi akuphatikizapo Linux Foundation, Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, ndi Palo Alto Networks. Mabungwe ena pafupifupi 40 alandiranso maitanidwe kuti achite nawo.

Yotulutsidwa mu February, chitsanzo cha Claude Opus 4.6 AI chinapeza magwiridwe antchito atsopano m'malo monga kuzindikira kufooka, kuzindikira ndi kukonza zolakwika, kuwunikanso kusintha, komanso kupanga ma code. Kuyesera ndi chitsanzo ichi cha AI kunalola kuzindikira kufooka kopitilira 500 m'mapulojekiti otseguka komanso kupanga compiler ya C yomwe imatha kupanga kernel ya Linux. Komabe, Claude Opus 4.6 sinachite bwino popanga ntchito zogwirira ntchito.

Malinga ndi Anthropic, chitsanzo cha m'badwo wotsatira cha "Claude Mythos" chimachita bwino kwambiri kuposa Claude Opus 4.6 popanga zinthu zokonzeka kugwiritsidwa ntchito. Mwa mazana angapo omwe anayesa kupanga zinthu zosayenera zomwe zapezeka mu injini ya JavaScript ya Firefox, ziwiri zokha ndi zomwe zinapambana ndi Claude Opus 4.6. Pobwereza kuyeseraku pogwiritsa ntchito mtundu woyambirira wa chitsanzo cha Mythos, zinthu zogwirira ntchito zinapangidwa nthawi 181—chiŵerengero cha kupambana chinawonjezeka kuchoka pa zero kufika pa 72.4%.

Anthropic adalengeza Claude Mythos, chitsanzo cha AI chomwe chingathe kupanga zinthu zogwirira ntchito.

Kuphatikiza apo, Claude Mythos imakulitsa kwambiri luso lake lozindikira kufooka ndi zolakwika. Izi, kuphatikiza ndi kuyenerera kwake pakupanga kugwiritsa ntchito, zimapangitsa kuti pakhale zoopsa zatsopano pamakampani: kugwiritsa ntchito njira zopewera kufooka kwa tsiku lopanda zingwe kumatha kupangidwa ndi anthu omwe si akatswiri pakangopita maola ochepa. Zadziwika kuti luso lozindikira ndi kugwiritsa ntchito kufooka kwa Mythos lafika pamlingo waukadaulo, ndipo pakadali pano ndi lopitirira luso la akatswiri odziwa bwino ntchito.

Popeza kutsegula mwayi wopanda malire ku chitsanzo cha AI chokhala ndi luso lotere kumafuna kukonzekera kwa makampani, poyamba adaganiza zotsegula mtundu woyamba kwa gulu losankhidwa la akatswiri kuti achite ntchito yozindikira kufooka ndi kukonza zinthu zofunika kwambiri pa mapulogalamu ndi mapulogalamu otseguka. Kuti athandizire ntchitoyi, ndalama zothandizira ma tokeni a $100 miliyoni zaperekedwa, ndipo $4 miliyoni ziperekedwa ku mabungwe omwe akuthandiza chitetezo cha mapulojekiti otseguka.

Mu benchmark ya CyberGym, yomwe imayesa kuthekera kwa ma model kuzindikira kufooka kwawo, model ya Mythos idapeza chigoli cha 83.1%, pomwe Opus 4.6 idapeza chigoli cha 66.6%. Mu mayeso a khalidwe la ma code, ma model adawonetsa magwiridwe antchito otsatirawa:

MayesoNthanoOposa 4.6 SWE-bench Pro 77.8% 53.4% ​​​​Terminal-Bench 2.0 82.0% 65.4% SWE-bench Multimodal 59% 27.1% SWE-bench Zilankhulo Zambiri 87.3% 77.8% SWE-bench Yatsimikizika 93.9% 80.8%

Pa nthawi yoyeserayi, Anthropic, pogwiritsa ntchito chitsanzo cha Mythos AI, adatha kuzindikira zofooka zikwi zingapo zomwe sizinali zodziwika kale (za masiku 0) m'masabata ochepa okha, zomwe zambiri mwa izo zidawerengedwa kuti ndi zofunika kwambiri. Pakati pawo, adapeza zofooka mu OpenBSD TCP stack zomwe sizinadziwike kwa zaka 27, zomwe zidalola kuti makina akutali azigwa. Adapezanso zofooka za zaka 16 pakukhazikitsa kwa FFmpeg codec ya H.264, komanso zofooka mu H.265 ndi av1 codecs, zomwe zimagwiritsidwa ntchito pokonza zomwe zapangidwa mwapadera.

Zofooka zingapo zinapezeka mu kernel ya Linux zomwe zingathandize wogwiritsa ntchito wopanda mwayi kupeza ufulu wa mizu. Kugwirizanitsa zofooka izi pamodzi kunalola kuti pakhale zogwiritsa ntchito zomwe zingapeze ufulu wa mizu potsegula masamba apadera mu msakatuli wapaintaneti. Zogwiritsa ntchito zinapangidwanso zomwe zinalola kugwiritsa ntchito ma code ndi ufulu wa mizu potumiza ma network packets opangidwa mwapadera ku seva ya FreeBSD NFS.

Kufooka kwapezeka mu makina ogwiritsira ntchito virtualization olembedwa m'chinenero chomwe chimapereka zida zotetezera zosungira kukumbukira. Kufooka kumeneku kumalola kugwiritsa ntchito ma code a mbali ya host kudzera mu kusintha kwa dongosolo la alendo (kufooka sikunatchulidwe chifukwa sikunakonzedwebe, koma kumawoneka kuti kulipo mu block yosatetezeka mu Rust code). Kufooka kwapezeka mu asakatuli onse otchuka a pa intaneti ndi malaibulale a cryptographic. Kufooka kwa SQL injection kwapezeka mu mapulogalamu osiyanasiyana apaintaneti.

Source: opennet.ru

Kuwonjezera ndemanga