麻豆高清

靸堧�滌毚 瓿店博 頂勲爤鞛勳泴韥� ‘SlotGCG’ 鞝滌晥… 鞓る姅 4鞗� ICLR 2026鞐愳劀 氚滍憸

鈻� (鞕检�诫秬韯�) 鞚挫毎歆� 甑愳垬, 鞝曥姽鞗� 靹濎偓瓿检爼(鞝�1鞝赌鞛�), 鞝曥��鞖� 靹濎偓瓿检爼, 旯�順勳�� 靹濎偓瓿检爼, 鞚挫湦靹� 靹濎偓瓿检爼

霃欔淡雽赌顷欔祼 旎错摠韯�·AI頃欕秬 鞚挫毎歆� 甑愳垬 鞐瓣惮韺�鞚� 瓯半寑鞏胳柎氇�雿�(LLM)鞚� 鞎堨爠靹� 旆�鞎届爯鞚� 鞝曤皜頃橁矊 攵勳劃頃� 鞐瓣惮 靹标臣毳� 氚滍憸頄堧嫟.

霃欔淡雽赌顷欔祼(齑濎灔 鞙れ灛鞗�)電� 氤戈祼 旎错摠韯�·AI頃欔臣 鞚戈车歆�電レ爠瓿� 鞝曥姽鞗�, 鞝曥��鞖�, 旯�順勳��, 鞚挫湦靹� 靹濎偓瓿检爼靸濌臣 鞚挫毎歆� 甑愳垬臧� 彀胳棳頃� 雲茧�� ‘SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks’臧� 鞓る姅 4鞗旍棎 臧滌禍霅橂姅 靹戈硠鞝� 甓岇渼鞚� 鞚戈车歆�電� 頃欗殞鞚� ICLR 2026(International Conference on Learning Representations)鞐� 毂勴儩霅橁碃 氚滍憸霅� 鞓堨爼鞚措澕瓿� 氚濏様雼�.

鞚措矆 鞐瓣惮電� 旮办〈 LLM 氤挫晥 鞖绊殞 瓿店博鞚� 頂勲‖頂勴姼 雭濍秬攵勳棎毵� 韱犿伆鞚� 於旉皜頃橂姅 氚╈嫕鞐� 歆戩�戫暣 鞕旊嫟電� 頃滉硠鞐愳劀 於滊皽頄堧嫟. 鞐瓣惮韺�鞚� 頂勲‖頂勴姼 雮措秬鞚� 韸轨爼 鞙勳箻鞐� 臁挫灛頃橂姅 旆�鞎� 歆�鞝愳潉 頇曥澑頃橁碃, 鞚措ゼ 鞝曤焿頇旐暅 Vulnerable Slot Score(VSS)鞕� 靸堧�滌毚 瓿店博 頂勲爤鞛勳泴韥� ‘SlotGCG’毳� 鞝滌晥頄堧嫟.

鞁ろ棙 瓴瓣臣, ‘SlotGCG’電� AdvBench 雿办澊韯办厠瓿� Llama, Mistral, Vicuna, Qwen 霌� 雼れ枒頃� 瓿店皽 LLM鞐愳劀 旮办〈 氚╈嫕氤措嫟 韽夑窢 鞎� 14% 雴掛潃 靹彪姤鞚� 氤挫榾瓿�, 斓滊寑 10氚� 牍犽ジ 靻嶋弰毳� 旮半�濏枅雼�. 霕� 氚╈柎 旮半矔鞚� 鞝侅毄霅� 頇橁步鞐愳劀霃� 旮办〈氤措嫟 29% 雴掛潃 瓿店博 靹标车毳犾潉 氤挫棳, 順勳灛 LLM 氚╈柎 觳搓硠鞚� 頃滉硠毳� 歃濍獏頄堧嫟.

鞐瓣惮韺�鞚� “鞚措矆 鞐瓣惮電� 雿旍毐 瓴�瓿犿暅 LLM 氚╈柎 旮半矔鞚� 靹り硠頃橁赴 鞙勴暅 AI 鞎堨爠靹� 瓴�歃� 氚� 霠堧摐韺� 鞐瓣惮霛茧姅 鞝愳棎靹� 鞚橂�戈皜 韥�雼�”氅� “雼�靾滍暅 瓿店博 旮半矔 鞝滌晥鞐� 攴胳箻歆� 鞎婈碃, 瓯半寑鞏胳柎氇�雿胳澊 鞏措枻 鞙勳箻 鞝曤炒鞐� 旆�鞎巾暅歆� 觳搓硠鞝侅溂搿� 鞚错暣頃橁碃 鞝侅毄頃� 靾� 鞛堧姅 鞐瓣惮”霛缄碃 靹る獏頄堧嫟. 韸鬼瀳, 鞚措矆 鞐瓣惮臧� 頂勲‖頂勴姼 雮� 鞙勳箻氤� 旆�鞎届劚鞚� 瓿勲焿鞝侅溂搿� 攵勳劃頄堧嫟電� 鞝愳棎靹�, 頄ロ泟 雽�攴滊�� 鞏胳柎氇�雿胳潣 鞎堨爠靹� 韽夑皜, 鞝曤牞(alignment) 旮办垹 瓿犽弰頇�, 氚╈柎 頂勲爤鞛勳泴韥� 靹り硠鞐� 韽�雱撽矊 頇滌毄霅� 瓴冹溂搿� 旮半寑霅滊嫟.
  

頃滍幐, 鞐瓣惮韺�鞚� 鞎烄溂搿� 雼れ枒頃� 氇�雿戈臣 鞁れ牅 靹滊箘鞀� 頇橁步鞐愳劀鞚� 鞚戈车歆�電� 旆�鞎届劚 韺岇晠瓿� 鞎堨爠頃� 鞚戈车歆�電� 臧滊皽鞚� 氇╉憸搿� 鞐瓣惮毳� 鞚挫柎臧� 瓿勴殟鞚措嫟.