\n<\/p>\n
\u00c7inli yapay zek\u00e2 giri\u015fimi Z.ai, bug\u00fcn yapt\u0131\u011f\u0131 a\u00e7\u0131klamada, a\u00e7\u0131k kaynakl\u0131 GLM-5.3 modelinin yaz\u0131l\u0131m a\u00e7\u0131klar\u0131n\u0131 tespit etme testlerinde Anthropic’in eri\u015fimi k\u0131s\u0131tl\u0131 Mythos 5 modeline yakla\u015ft\u0131\u011f\u0131n\u0131 bildirdi.<\/p>\n
\u00a0<\/p>\n
Bu geli\u015fme, Bat\u0131l\u0131 geli\u015ftiriciler aras\u0131nda ilgi g\u00f6ren \u00c7inli yapay zek\u00e2 giri\u015fiminin konumunu peki\u015ftirdi.<\/p>\n
\u00a0<\/p>\n
Z.ai, bir modelin kodlar\u0131 inceleme, g\u00fcvenlik a\u00e7\u0131klar\u0131n\u0131 belirleme ve bunlar\u0131n ger\u00e7ekli\u011fini do\u011frulama kabiliyetini \u00f6l\u00e7en CyberGym testinde GLM-5.3’\u00fcn y\u00fczde 84,5 puan ald\u0131\u011f\u0131n\u0131 a\u00e7\u0131klad\u0131.<\/p>\n
\u00a0<\/p>\n
\u015eirket, Mythos 5 i\u00e7in bu oran\u0131n y\u00fczde 83,8 oldu\u011funu bildirdi. Bu sonu\u00e7lar ba\u011f\u0131ms\u0131z kaynaklarca teyit edilmedi.<\/p>\n
\u00a0<\/p>\n
GLM-5.3, savunma ama\u00e7l\u0131 g\u00fcvenlik ara\u015ft\u0131rmalar\u0131n\u0131n standart bir par\u00e7as\u0131 olan a\u00e7\u0131klar\u0131 \u00e7al\u0131\u015fan sald\u0131r\u0131lara d\u00f6n\u00fc\u015ft\u00fcrme a\u015famas\u0131nda Mythos 5’in gerisinde kald\u0131.<\/p>\n
\u00a0<\/p>\n
Z.ai, bu kabiliyeti \u00f6l\u00e7en ExploitBench testinde kendi modelinin y\u00fczde 54,4, Mythos 5’in ise y\u00fczde 78,0 puan ald\u0131\u011f\u0131n\u0131 belirtti.<\/p>\n
\u00a0<\/p>\n
Zamana dayal\u0131 ayr\u0131 bir testte Z.ai, GLM-5.3’\u00fcn iki saatte 105, alt\u0131 saatte ise 130 sald\u0131r\u0131 geli\u015ftirme g\u00f6revini tamamlad\u0131\u011f\u0131n\u0131 a\u00e7\u0131klad\u0131.\u00a0<\/p>\n
\u00a0<\/p>\n
Mythos 5 ise ayn\u0131 s\u00fcrelerde s\u0131ras\u0131yla 181 ve 247 g\u00f6revi tamamlad\u0131.<\/p>\n
\u00a0<\/p>\n
Anthropic, siber g\u00fcvenlik