{"id":11417,"date":"2025-11-30T21:30:24","date_gmt":"2025-11-30T12:30:24","guid":{"rendered":"https:\/\/www.stechstar.com\/user\/wordpress\/?p=11417"},"modified":"2026-08-20T03:16:10","modified_gmt":"2026-08-19T18:16:10","slug":"%ec%9d%b8%ea%b3%b5%ec%a7%80%eb%8a%a5-%ec%84%9c%eb%b9%84%ec%8a%a4-%eb%b3%b4%ec%95%88-%ed%83%88%ec%98%a5-%ea%b3%b5%ea%b2%a9%ec%9c%bc%eb%a1%9c%eb%b6%80%ed%84%b0-llm%ec%9d%84-%eb%b3%b4%ed%98%b8%ed%95%98","status":"publish","type":"post","link":"https:\/\/www.stechstar.com\/user\/wordpress\/%ec%9d%b8%ea%b3%b5%ec%a7%80%eb%8a%a5-%ec%84%9c%eb%b9%84%ec%8a%a4-%eb%b3%b4%ec%95%88-%ed%83%88%ec%98%a5-%ea%b3%b5%ea%b2%a9%ec%9c%bc%eb%a1%9c%eb%b6%80%ed%84%b0-llm%ec%9d%84-%eb%b3%b4%ed%98%b8%ed%95%98\/","title":{"rendered":"[\uc778\uacf5\uc9c0\ub2a5 \uc11c\ube44\uc2a4 \ubcf4\uc548] \ud0c8\uc625 \uacf5\uaca9\uc73c\ub85c\ubd80\ud130 LLM\uc744 \ubcf4\ud638\ud558\ub294 \ubc29\ubc95 , How to Protect LLMs from Jailbreaking Attacks"},"content":{"rendered":"<h1>[\uc778\uacf5\uc9c0\ub2a5 \uc11c\ube44\uc2a4 \ubcf4\uc548] \ud0c8\uc625 \uacf5\uaca9\uc73c\ub85c\ubd80\ud130 LLM\uc744 \ubcf4\ud638\ud558\ub294 \ubc29\ubc95 , How to Protect LLMs from Jailbreaking Attacks<\/h1>\n<div id=\"content\" data-events=\"scroll\" data-magellan=\"rz7y9r-magellan\" data-resize=\"content\" data-scroll=\"content\">\n<div class=\"blog-hero-wrapper\" data-aa-hero=\"blog\">\n<div class=\"row\">\n<div class=\"columns large-9 medium-12 small-12\">\n<div class=\"hero-container blog-hero\">\n<div class=\"hero-data blog-hero\">\n<h1><span dir=\"auto\">\ud0c8\uc625 \uacf5\uaca9\uc73c\ub85c\ubd80\ud130 LLM\uc744 \ubcf4\ud638\ud558\ub294 \ubc29\ubc95<\/span><\/h1>\n<div data-aa-comp-name=\"blog-hero:author-link\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">\ub178\uc544 \ud50c\ub77c\uc774\uc288\ub9cc, \uc5d0\uc774\ubbf8 \uc640\uace0\ub108 \ubc15\uc0ac, \uc559\ub4dc\ub808 \uc751\uc6b0\uc60c \ubc15\uc0ac<\/span><\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div class=\"blog-top-container\">\n<div class=\"generic-block\">\n<div class=\"row\">\n<div class=\"small-12 columns\">\n<div class=\"image parbase\">\n<div class=\"image-container \n\n        no-pointer\n        default-style \n        default-style \n        default-style\n\n        \"><img decoding=\"async\" class=\"image widecq-dd-image\" src=\"https:\/\/www.stechstar.com\/user\/zbxe\/files\/attach\/images\/112899\/915\/112\/70eedf6e23c9b9e5c209b083506e56dc.jpg\" alt=\"\" data-autoattach=\"success\" data-emptytext=\"Image\" data-cke-saved-src=\".\/files\/attach\/images\/112899\/915\/112\/70eedf6e23c9b9e5c209b083506e56dc.jpg\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div class=\"sticky-sidebar-container full\">\n<div id=\"sticky-anchor\">\n<div class=\"row\">\n<div class=\"blog-breakout-parsys small-12 large-10 large-offset-1 columns\">\n<div>&nbsp;<\/div>\n<\/div>\n<\/div>\n<div class=\"row\">\n<div class=\"grid-column--center-align columns large-10 full\">\n<div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1628759649&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h2 id=\"booz-allen-examines-how-to-strengthen-model-integrity\" class=\"title title-component title-component--default align-left\"><span dir=\"auto\">Booz Allen\uc740 \ubaa8\ub378 \ubb34\uacb0\uc131\uc744 \uac15\ud654\ud558\ub294 \ubc29\ubc95\uc744 \uc870\uc0ac\ud569\ub2c8\ub2e4.<\/span><\/h2>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span class=\"rte-large\"><span dir=\"auto\">\uc5f0\ubc29 \uae30\uad00\ub4e4\uc740 \uc5c5\ubb34\ub97c \uac04\uc18c\ud654\ud558\uace0 \uc9c8\ubb38\uc5d0 \ub2f5\ud558\uae30 \uc704\ud574 Llama-2 \ubc0f ChatGPT\uc640 \uac19\uc740 \ub300\uaddc\ubaa8 \uc5b8\uc5b4 \ubaa8\ub378(LLM)\uc744 \uc6b4\uc601\uc5d0 \uc810\uc810 \ub354 \ud1b5\ud569\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc5d4\uc9c0\ub2c8\uc5b4\ub4e4\uc740 \uc774\ub7ec\ud55c \ubaa8\ub378\uc774 &#8220;\uc720\uc6a9\ud558\uace0 \ubb34\ud574\ud558\ub3c4\ub85d&#8221; \uc124\uacc4\ud558\uace0 \uc704\ud5d8\ud55c \uc694\uccad\uc740 \uac70\ubd80\ud569\ub2c8\ub2e4. \ubbf8\uc138 \uc870\uc815, \uc778\uac04 \ud53c\ub4dc\ubc31\uc744 \ud65c\uc6a9\ud55c \uac15\ud654 \ud559\uc2b5, \uc9c1\uc811 \uc120\ud638\ub3c4 \ucd5c\uc801\ud654\uc640 \uac19\uc740 \uae30\ubc95\uc740 \ubaa8\ub378 \uc548\uc804\uc131\uc744 \ub354\uc6b1 \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \uc774\ub7ec\ud55c \uc870\uce58\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0, \ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\ub77c\ub294 \uc2ec\uac01\ud55c LLM \ucde8\uc57d\uc810\uc774 AI \uc2dc\uc2a4\ud15c\uc744 \uacc4\uc18d \uc704\ud5d8\uc5d0 \ube60\ub728\ub9ac\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/span><\/p>\n<p><span dir=\"auto\">\ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\ub294 LLM\uc774 \ud574\uc11c\ub294 \uc548 \ub420 \uc77c\uc744 \ud558\ub3c4\ub85d \uc18d\uc774\ub294 \ud2b9\uc815 \uc785\ub825\uac12\uc785\ub2c8\ub2e4. \uad50\ubb18\ud558\uac8c \uc124\uacc4\ub418\uc5c8\uc9c0\ub9cc \uc545\uc758\uc801\uc778 \uc774\ub7ec\ud55c \ud504\ub86c\ud504\ud2b8\ub294 \uac00\uc7a5 \uac15\ub825\ud55c \ubcf4\uc548 \uc870\uce58\uc870\ucc28 \uc6b0\ud68c\ud558\uc5ec \uc5f0\ubc29 \uc815\ubd80 \uc5c5\ubb34\uc5d0 \uc2ec\uac01\ud55c \uc704\ud5d8\uc744 \ucd08\ub798\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\uc774\ub7ec\ud55c \uacfc\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \ubd80\uc988 \uc568\ub7f0\uc740 \ud0c8\uc625\uc5d0 \ub300\ud55c \uc0c8\ub85c\uc6b4 \ubc29\uc5b4\ucc45\uc744 \ubaa8\uc0c9\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc811\uadfc \ubc29\uc2dd\uc740 \uae30\uad00\uc5d0 \uc0c1\ub2f9\ud55c \uc784\ubb34 \uc774\uc810\uc744 \uc81c\uacf5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc989, LLM\uc774 \ubb34\ud574\ud55c \ud504\ub86c\ud504\ud2b8\uc5d0 \uc751\ub2f5\ud558\ub294 \ub2a5\ub825\uc744 \uc800\ud574\ud558\uc9c0 \uc54a\uc73c\uba74\uc11c LLM\uc744 \ubcf4\ud638\ud558\uc5ec \uae30\uc5c5 \uc0dd\uc0b0\uc131 \ud5a5\uc0c1\uc758 \uc6d0\ub3d9\ub825\uc73c\ub85c \uacc4\uc18d \uae30\ub2a5\ud560 \uc218 \uc788\ub3c4\ub85d \ud569\ub2c8\ub2e4.<\/span><\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1403034696&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"how-jailbreaking-works\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"how-jailbreaking-works\"><span dir=\"auto\">\ud0c8\uc625 \uc791\ub3d9 \ubc29\uc2dd<\/span><\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">\uacf5\uaca9\uc790\ub294 \ud0c8\uc625 \uba54\uc2dc\uc9c0\ub97c \uc870\uc791\ud558\uc5ec LLM\uc744 \uc870\uc791\ud558\uc5ec \uac1c\uc778 \uc2dd\ubcc4 \uc815\ubcf4\uc640 \uac19\uc740 \ubbfc\uac10\ud55c \uc815\ubcf4\ub97c \ub178\ucd9c\ud558\uac70\ub098, \ubd88\ubc95 \ud589\uc704(\uc608: \ud3ed\ud0c4 \uc81c\uc870, \ud53c\uc2f1 \uc774\uba54\uc77c \uc791\uc131) \ub610\ub294 \uc99d\uc624 \ud45c\ud604\uacfc \uac19\uc740 \uc720\ud574\ud55c \ucf58\ud150\uce20\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4. \uacf5\uaca9\uc790\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ub2e4\uc591\ud558\uace0 \ucc3d\uc758\uc801\uc778 \uae30\ubc95\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/span><\/p>\n<ul>\n<li><span dir=\"auto\">\ub864 \ud50c\ub808\uc774:<\/span><span dir=\"auto\">&nbsp;\ubbfc\uac10\ud558\uac70\ub098 \ubcf4\ud638\ub41c \uc815\ubcf4\ub97c \uacf5\uac1c\ud558\ub294 \ub370 \ub300\ud55c \uae30\uc874 \uac00\ub4dc\ub808\uc77c\uc744 \uc6b0\ud68c\ud558\uae30 \uc704\ud574 AI\uc5d0\uac8c \ud2b9\uc815 \uc5ed\ud560\uc774\ub098 \ud398\ub974\uc18c\ub098\ub97c \uc218\ud589\ud558\ub3c4\ub85d \uc694\uccad\ud558\ub294 \uac83\uc785\ub2c8\ub2e4.<\/span><\/li>\n<li><span dir=\"auto\">\uc8fc\uc758 \uc804\ud658:<\/span><span dir=\"auto\">&nbsp;AI\uc758 \uc548\uc804 \ud504\ub85c\ud1a0\ucf5c\uc744 \uc6b0\ud68c\ud558\uae30 \uc704\ud574 AI\uc758 \uc8fc\uc758\ub97c \ubd84\uc0b0\uc2dc\ud0a4\ub294 \ud589\uc704.<\/span><\/li>\n<li><span dir=\"auto\">\uad8c\ud55c \ud655\ub300:<\/span><span dir=\"auto\">&nbsp;AI \uc2dc\uc2a4\ud15c \ub0b4\uc5d0\uc11c \ub354 \ub192\uc740 \uc218\uc900\uc758 \uc561\uc138\uc2a4 \uad8c\ud55c\uc744 \uc5bb\ub294 \uac83\uc785\ub2c8\ub2e4.<\/span><\/li>\n<li><span dir=\"auto\">\uc811\ub450\uc0ac \uc0bd\uc785:<\/span><span dir=\"auto\">&nbsp;AI\uc758 \uc751\ub2f5\uc744 \uc870\uc791\ud558\uae30 \uc704\ud574 \ud2b9\uc815 \uad6c\ubb38\uc744 \ucd94\uac00\ud569\ub2c8\ub2e4.<\/span><\/li>\n<li><span dir=\"auto\">\uac70\ubd80 \uc5b5\uc81c:<\/span><span dir=\"auto\">&nbsp;AI\uac00 \uac70\ubd80 \ud504\ub85c\ud1a0\ucf5c\uc744 \ubb34\uc2dc\ud558\ub3c4\ub85d \uc18d\uc774\ub294 \uac83\uc785\ub2c8\ub2e4.<\/span><\/li>\n<li><span dir=\"auto\">\ub2e8\uc5b4 \uac8c\uc784\uacfc \ub09c\ub3c5\ud654:<\/span><span dir=\"auto\">&nbsp;\ubcf5\uc7a1\ud55c \uc5b8\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc5ec AI\ub97c \ud63c\ub780\uc2a4\ub7fd\uac8c \ud568.<\/span><\/li>\n<li><span dir=\"auto\">\ub2e4\uad6d\uc5b4 \uc785\ub825 \ubc0f \uc554\ud638:<\/span><span dir=\"auto\">&nbsp;\ubcf4\uc548\uc744 \uc6b0\ud68c\ud558\uae30 \uc704\ud574 \ub2e4\uc591\ud55c \uc5b8\uc5b4\ub098 \uc554\ud638\ud654\ub41c \uba54\uc2dc\uc9c0\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/span><\/li>\n<\/ul>\n<p><span dir=\"auto\">\uc774\ub7ec\ud55c \uae30\ubc95\uc740 \uc5bc\ub9c8\ub098 \ud6a8\uacfc\uc801\uc77c\uae4c\uc694? \uc5f0\uad6c\uc5d0 \ub530\ub974\uba74&nbsp; \ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\ub294&nbsp;Vicuna, ChatGLM3, GPT-3.5, PaLM2\ub97c \ud3ec\ud568\ud55c \ub2e4\uc591\ud55c LLM\uc5d0\uc11c&nbsp;<\/span><a title=\"\ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\" href=\"https:\/\/arxiv.org\/abs\/2402.05668\" target=\"_blank\" rel=\"noopener\" data-cke-saved-href=\"https:\/\/arxiv.org\/abs\/2402.05668\"><span dir=\"auto\">\uc9c0\uc18d\uc801\uc73c\ub85c \ub192\uc740 \uacf5\uaca9 \uc131\uacf5\ub960\uc744 \ubcf4\uc774\uba70 , \uc774\ub294 \uc548\uc804 \uc870\uce58\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 \uadf8 \uacac\uace0\uc131\uacfc \uc804\uc774\uc131\uc744 \uac15\uc870\ud569\ub2c8\ub2e4. \ub610\ud55c,&nbsp;<\/span><\/a><a title=\"\uc790\ub3d9\ud654\ub41c \ub098\ubb34\" href=\"https:\/\/arxiv.org\/abs\/2312.02119\" target=\"_blank\" rel=\"noopener\" data-cke-saved-href=\"https:\/\/arxiv.org\/abs\/2312.02119\"><span dir=\"auto\">\uc790\ub3d9\ud654\ub41c \ud2b8\ub9ac \uc624\ube0c \uc5b4\ud0dd\uc2a4 \ud504\ub8e8\ub2dd<\/span><\/a><span dir=\"auto\">&nbsp;(TAP) \uae30\ubc95\uc740 \ub2e8 \uba87 \ubc88\uc758 \ucffc\ub9ac\ub9cc\uc73c\ub85c 80% \uc774\uc0c1\uc758 \uc131\uc219 LLM \ud504\ub86c\ud504\ud2b8\ub97c \uc131\uacf5\uc801\uc73c\ub85c \ud0c8\uc625\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\ud0c8\uc625\uc774 \uc65c \uadf8\ub807\uac8c \ud6a8\uacfc\uc801\uc77c\uae4c\uc694? \uc774\ub294 LLM \ub0b4\ubd80\uc758 \uc0c1\ucda9\ub418\ub294 \ubaa9\ud45c\uc640 \uad00\ub828\uc774 \uc788\uc2b5\ub2c8\ub2e4. \ud55c\ud3b8\uc73c\ub85c\ub294 \ubaa8\ub378\uc774 \uc9c0\uce68\uc744 \uc900\uc218\ud558\uc5ec \ub3c4\uc6c0\uc744 \uc8fc\ub294 \uac83\uc744 \ubaa9\ud45c\ub85c \ud558\uc9c0\ub9cc, \ub2e4\ub978 \ud55c\ud3b8\uc73c\ub85c\ub294 \ud53c\ud574\ub97c \uc785\ud788\uc9c0 \uc54a\uae30 \uc704\ud574 \uc548\uc804 \uc9c0\uce68\uc744 \uc900\uc218\ud574\uc57c \ud569\ub2c8\ub2e4. \ub2e4\ub978 \uacbd\uc6b0\uc5d0\ub294 \uc77c\ubc18\ud654\uc758 \ubd88\uc77c\uce58 \ub54c\ubb38\uc77c \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc989, \uc548\uc804 \ud6c8\ub828 \ub370\uc774\ud130\uc758 \uc591\uc774 \uc804\uccb4 \ud6c8\ub828 \ub370\uc774\ud130\ub97c \uc644\uc804\ud788 \ub300\ud45c\ud558\uc9c0 \ubabb\ud558\uc5ec \uc77c\ubd80 \uacf5\uaca9 \uc2dc\ub098\ub9ac\uc624\uac00 \ub204\ub77d\ub418\ub294 \uac83\uc785\ub2c8\ub2e4. \uc624\ub298\ub0a0 \uc6b0\ub9ac\ub294 \uc0c8\ub85c\uc6b4 \ud0c8\uc625 \uacf5\uaca9\uc774 \ub4f1\uc7a5\ud558\uace0, \ub05d\uc5c6\ub294 \uace0\uc591\uc774\uc640 \uc950\uc758 \uac8c\uc784\uc5d0\uc11c \uc0c8\ub85c\uc6b4 \ubc29\uc5b4\ucc45\uc774 \uac1c\ubc1c\ub418\ub294 \uc545\uc21c\ud658\uc744 \ubaa9\uaca9\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-78211345&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"risk-scenarios-for-government\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"risk-scenarios-for-government\"><span dir=\"auto\">\uc815\ubd80\ub97c \uc704\ud55c \uc704\ud5d8 \uc2dc\ub098\ub9ac\uc624<\/span><\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">LLM\uc5d0 \ub300\ud55c \ud0c8\uc625 \uacf5\uaca9\uc740 \uc5f0\ubc29 \uae30\uad00\uc5d0 \uc2ec\uac01\ud55c \uc704\ud5d8\uc744 \ucd08\ub798\ud569\ub2c8\ub2e4. \uad6d\uac00 \uc548\ubcf4\uc640 \uad00\ub828\ub41c \uc704\ud5d8\uc5d0\ub294 \ub370\uc774\ud130 \uc720\ucd9c, \uac1c\uc778\uc815\ubcf4 \ubcf4\ud638 \uce68\ud574, \ud5c8\uc704 \uc815\ubcf4 \ud655\uc0b0, \uc790\ub3d9\ud654 \uc2dc\uc2a4\ud15c \uc870\uc791, \uadf8\ub9ac\uace0 \uc758\uc0ac \uacb0\uc815 \ud504\ub85c\uc138\uc2a4 \uce68\ud574 \ub4f1\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\ud0c8\uc625\ub41c LLM\uc774 \ub300\ud53c \uc808\ucc28\uc5d0 \ub300\ud55c \ud5c8\uc704 \uc815\ubcf4\ub97c \uba54\uc2dc\uc9c0\uc5d0 \uc785\ub825\ud558\uc5ec \ud63c\ub780\uc744 \uc57c\uae30\ud558\uace0 \uc0dd\uba85\uc744 \uc704\ud611\ud558\ub294 \uad6d\uac00 \ube44\uc0c1 \uc0c1\ud669\uc744 \uc0c1\uc0c1\ud574 \ubcf4\uc138\uc694. \ub610\ub294 \uacf5\uaca9\uc790\uac00 \ud0c8\uc625\ub41c LLM\uc744 \uc0ac\uc6a9\ud558\uc5ec \uac10\uc2dc \uc81c\uc5b4 \ubc0f \ub370\uc774\ud130 \uc218\uc9d1 \uc2dc\uc2a4\ud15c\uc5d0 \uc545\uc131 \uba85\ub839\uc744 \uc8fc\uc785\ud558\uc5ec \ub300\ub3c4\uc2dc\uc5d0 \uc815\uc804\uc744 \uc720\ubc1c\ud558\ub294 \uc0c1\ud669\uc744 \uc0c1\uc0c1\ud574 \ubcf4\uc138\uc694.<\/span><\/p>\n<p><span dir=\"auto\">\ud0c8\uc625\ub41c LLM\uc740 \uc724\ub9ac\uc801, \ubc95\uc801 \ubb38\uc81c\ub97c \uc57c\uae30\ud560 \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4. \uc190\uc0c1\ub41c \ubaa8\ub378\uc740 \ubd88\ubc95\uc801\uc778 \ucf58\ud150\uce20\ub97c \uc0dd\uc131\ud558\uc5ec \uae30\uad00\uc774 \uaddc\uc81c \ucc98\ubc8c\uc744 \ubc1b\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub610\ud55c \ud0c8\uc625\ub41c LLM\uc774 \ucc28\ubcc4\uc801\uc774\uac70\ub098 \ubd88\ucf8c\uac10\uc744 \uc8fc\ub294 \ucf58\ud150\uce20\ub97c \uc0dd\uc131\ud558\uc5ec \uacf5\uc2dd \ubcf4\uace0\uc11c\uc5d0 \ud3ec\ud568\ub418\uba74 \ud574\ub2f9 \uae30\uad00\uc740 \ud3c9\ud310\uc774 \uc190\uc0c1\ub418\uace0 \uc18c\uc1a1\uc5d0 \uc9c1\uba74\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.&nbsp;<\/span><\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1350676556&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"attack-types\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"attack-types\"><span dir=\"auto\">\uacf5\uaca9 \uc720\ud615<\/span><\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">\ucd5c\uadfc\uc758 \ubc1c\uc804\uc73c\ub85c \uc785\ub825 \ud14d\uc2a4\ud2b8\uc758 \uc77c\ubd80\ub97c \uc218\uc815\ud558\uc5ec \ub354\uc6b1 \uac15\ub825\ud55c \ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\uac00 \uac00\ub2a5\ud574\uc84c\uc2b5\ub2c8\ub2e4. \ubcf4\uc548 \uc804\ubb38\uac00\ub4e4\uc740 \uc774 \uacfc\uc815\uc744 &#8220;\uc12d\ub3d9(perturbation)&#8221;\uc774\ub77c\uace0 \ubd80\ub985\ub2c8\ub2e4. \uc0dd\uc131 AI \uc678\uc5d0\ub3c4 \uc5f0\uad6c\uc790\ub4e4\uc740 \uc774\ub7ec\ud55c \ubcc0\ud654\uac00 \uac10\uc815 \ubd84\uc11d \ubc0f \uc720\ud574\uc131 \ud0d0\uc9c0\uc640 \uac19\uc740 \ud14d\uc2a4\ud2b8 \ubd84\ub958 \ubb38\uc81c\uc5d0 \uc5b4\ub5a4 \uc601\ud5a5\uc744 \ubbf8\uce58\ub294\uc9c0 \uc5f0\uad6c\ud574 \uc654\uc2b5\ub2c8\ub2e4. \uc720\ud574\ud55c \uc785\ub825 \ubc0f \ucd9c\ub825 \ud0d0\uc9c0\ub294 \ub9e4\uc6b0 \uc911\uc694\ud558\ubbc0\ub85c, \uc0dd\uc131 \ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\uc640 \ud568\uaed8 \uc0ac\uc6a9\ub418\ub294 \uc801\ub300\uc801 \ubd84\ub958 \uacf5\uaca9\uc744 \uace0\ub824\ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4. \uc12d\ub3d9 \uae30\ubc18 \uacf5\uaca9\uc758 \uc608\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<ul>\n<li><span dir=\"auto\">\ubb38\uc790 \uc218\uc900 \uad50\ub780(CLP):<\/span><span dir=\"auto\">&nbsp;\uacf5\uaca9\uc790\ub294 \ud504\ub86c\ud504\ud2b8 \ubb38\uc790\ub97c \ubb34\uc791\uc704\ub85c \uad50\uccb4, \uc0bd\uc785 \ub610\ub294 \uc0ad\uc81c\ud558\uc5ec \uc911\uc694 \ub2e8\uc5b4\ub97c \ud45c\uc801\uc73c\ub85c \uc0bc\uc544 \ub9e4\ud551\uc744 \ubcc0\uacbd\ud569\ub2c8\ub2e4. \uc774\ub807\uac8c \ud558\uba74 \ubaa8\ub378\uc774 \uc720\ud574\ud55c \ud504\ub86c\ud504\ud2b8\ub97c \uac10\uc9c0\ud558\ub294 \ub370 \uc0ac\uc6a9\ud558\ub294 \uc815\ubcf4\uac00 \uc81c\uac70\ub429\ub2c8\ub2e4. CLP\ub294 \uc624\ud0c0\ub85c \uc624\uc778\ub420 \uc218\ub3c4 \uc788\uace0, \uc0ac\uc2e4\uc0c1 \ub208\uc5d0 \ub744\uc9c0 \uc54a\uc744 \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4(\uc608: \ub3d9\ud615\uc774\uc758\uc5b4).<\/span><\/li>\n<\/ul>\n<ul>\n<li><span dir=\"auto\">\ub2e8\uc5b4 \uc218\uc900\uc758 \uad50\ub780(WLP):<\/span><span dir=\"auto\">&nbsp;\uacf5\uaca9\uc790\ub294 \uc911\uc694\ud55c \ub2e8\uc5b4\ub97c \ub3d9\uc758\uc5b4\ub098 \uc5b4\ud718\uc5d0 \uc5c6\ub294 \ub2e8\uc5b4\ub85c \ub300\uccb4\ud558\uc5ec \ubaa8\ub378\uc758 \ub3d9\uc791\uc744 \ud06c\uac8c \ubcc0\uacbd\ud569\ub2c8\ub2e4.<\/span><\/li>\n<\/ul>\n<ul>\n<li><span dir=\"auto\">\ubb38\uc7a5 \uc218\uc900 \uad50\ub780(SLP):<\/span><span dir=\"auto\">&nbsp;\uacf5\uaca9\uc790\ub294 \ud504\ub86c\ud504\ud2b8\uc758 \uc6d0\ub798 \uc758\ubbf8\ub294 \uc720\uc9c0\ud558\ub418 \ubaa8\ub378\uc744 \ud63c\ub780\uc2a4\ub7fd\uac8c \ud558\uae30 \uc704\ud574 \ud504\ub86c\ud504\ud2b8\ub97c \ubc14\uafd4 \ud45c\ud604\ud558\uac70\ub098 \ubcc0\uacbd\ud569\ub2c8\ub2e4. \uad00\ub828 \uc5c6\ub294 \ubb38\uc7a5\uc744 \ucd94\uac00\ud558\uac70\ub098 \ud55c \uc5b8\uc5b4\uc5d0\uc11c \ub2e4\ub978 \uc5b8\uc5b4\ub85c, \uadf8\ub9ac\uace0 \uadf8 \ubc18\ub300\ub85c &#8220;\uc655\ubcf5&#8221; \ubc88\uc5ed\uc744 \uc0ac\uc6a9\ud558\ub294 \uac83\uc774 \uc77c\ubc18\uc801\uc778 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/span><\/li>\n<\/ul>\n<ul>\n<li><span dir=\"auto\">\ud0d0\uc695\uc801 \uc88c\ud45c \uae30\uc6b8\uae30(GCG) \uacf5\uaca9:<\/span><span dir=\"auto\">&nbsp;GCG \uacf5\uaca9\uc740 \ubaa8\ub378\uc758 \uc548\uc804 \uae30\ub2a5\uc744 \uc6b0\ud68c\ud558\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub418\ub294 \uc811\ubbf8\uc0ac(\uac89\ubcf4\uae30\uc5d0 \ubb34\uc791\uc704\uc801\uc778 \ubb38\uc790\uc640 \ub2e8\uc5b4\uc758 \ubb38\uc790\uc5f4)\ub97c \uc0dd\uc131\ud558\uc5ec \uc720\ud574\ud55c \ud504\ub86c\ud504\ud2b8\uac00 \ud0c0\uac9f\uc73c\ub85c \ud558\ub294 \uae0d\uc815\uc801 \uc751\ub2f5\uc744 \uc0dd\uc131\ud560 \uac00\ub2a5\uc131\uc744 \uadf9\ub300\ud654\ud569\ub2c8\ub2e4.<\/span><\/li>\n<\/ul>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-6844399&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"defenses-against-jailbreaking\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"defenses-against-jailbreaking\"><span dir=\"auto\">\ud0c8\uc625\uc5d0 \ub300\ud55c \ubc29\uc5b4\ucc45<\/span><\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">\ud0c8\uc625\uc744 \ud3c9\uac00\ud558\ub294 \uac83\uc740 \ub9e4\uc6b0 \uc5b4\ub835\uace0, \uc5b4\ub5a4 \uc2dc\uc2a4\ud15c\uc5d0\uc11c\ub4e0 \uc624\ub958(\uac70\uc9d3 \uc591\uc131 \ubc0f \uac70\uc9d3 \uc74c\uc131)\ub294 \ubc1c\uc0dd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcf4\uc870 \ubaa8\ub378 \uc0ac\uc6a9\uacfc \uac19\uc740 \ub300\uc548\uc801\uc778 \ubc29\ubc95\uc740 \ubb38\uc790\uc5f4 \ub9e4\uce6d\uc5d0 \ube44\ud574 \uac70\uc9d3 \uc591\uc131 \ubc0f \uac70\uc9d3 \uc74c\uc131\ub960\uc744 \ub0ae\ucd9c \uc218 \uc788\uc9c0\ub9cc, \ubaa8\ub378\uc774 \ud504\ub86c\ud504\ud2b8\ub97c \uc774\ud574\ud558\uc9c0 \ubabb\ud558\uac70\ub098 \ubd80\ubd84\uc801 \uac70\ubd80\/\uc900\uc218\ub97c \ubc18\ud658\ud558\ub294 \uc0d8\ud50c\uc744 \ubd84\ub958\ud558\ub294 \uac83\uc5d0 \ub300\ud55c \uc758\ubb38\uc740 \uc5ec\uc804\ud788 \u200b\u200b\ub0a8\uc544 \uc788\uc2b5\ub2c8\ub2e4. &#8220;\ubd80\ubd84 \uac70\ubd80&#8221;\uc640 &#8220;\ubd80\ubd84\uc801 \uc900\uc218&#8221;\ub77c\ub294 \uc6a9\uc5b4\uc870\ucc28\ub3c4 \uc5f0\uad6c\uc790\ub4e4 \uc0ac\uc774\uc5d0\uc11c \ucd9c\ub825\uc774 \uc720\ud574\ud574\uc9c0\ub294 \uc2dc\uc810\uc5d0 \ub300\ud55c \uc758\uacac \ucc28\uc774\uac00 \uc788\uc744 \uc218 \uc788\uae30 \ub54c\ubb38\uc5d0 \ub17c\ub780\uc758 \uc5ec\uc9c0\uac00 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\ud0c8\uc625 \uacf5\uaca9\uc5d0 \ub300\ud55c \uac15\ub825\ud55c \ubc29\uc5b4 \uccb4\uacc4\ub97c \uad6c\ucd95\ud558\ub294 \uac83 \ub610\ud55c \ub2e4\uc591\ud55c \ubaa8\ub378\uacfc \uacf5\uaca9 \uc720\ud615\uc73c\ub85c \uc778\ud574 \uc5b4\ub824\uc6b4 \uac83\uc73c\ub85c \ub098\ud0c0\ub0ac\uc2b5\ub2c8\ub2e4. \ud604\uc7ac \ub9ce\uc740 \ubc29\uc5b4 \uccb4\uacc4\uac00 \ud2b9\uc815 \uc720\ud615\uc758 \ud0c8\uc625\uc744 \ud45c\uc801\uc73c\ub85c \uc0bc\uace0 \uc788\uc73c\uba70, \uacf5\uaca9\uacfc \uc720\uc0ac\ud55c \uad50\ub780 \uc804\ub7b5\uc744 \uc0ac\uc6a9\ud558\ub294 \uacbd\uc6b0\uac00 \ub9ce\uc2b5\ub2c8\ub2e4. \uc77c\ubd80 \ubc29\uc5b4 \uccb4\uacc4\ub294 \ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\uc758 \uc601\ud5a5\uc744 \uc904\uc774\uae30 \uc704\ud574 WLP(Walk-Loop)\uc640 SLP(Single-Loop)\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \ub610 \ub2e4\ub978 \uc811\uadfc\ubc95\uc740 CLP\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc801\ub300\uc801 \uc811\ubbf8\uc0ac\uc758 \ubb38\uc790\ub97c \ubb34\uc791\uc704\ub85c \ubcc0\uacbd\ud568\uc73c\ub85c\uc368 GCG \uc811\ubbf8\uc0ac\uc758 \uc601\ud5a5\uc744 \ucd5c\uc18c\ud654\ud569\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\uadf8\ub7ec\ub098 \uc774 \ubc29\ubc95\uc740 \uc6d0\ub798 \ud504\ub86c\ud504\ud2b8\uc640 \uc811\ubbf8\uc0ac\ub97c \uc790\ub3d9\uc73c\ub85c \uad6c\ubd84\ud558\ub294 \ub370 \uc5b4\ub824\uc6c0\uc744 \uacaa\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4 SmoothLLM \uc54c\uace0\ub9ac\uc998\uc740 \uc811\ubbf8\uc0ac\ub9cc \ubd84\ub9ac\ud558\ub294 \ub300\uc2e0 \uc811\ubbf8\uc0ac\ub97c \ud3ec\ud568\ud55c \uc804\uccb4 \uc785\ub825\uc758 \uc77c\uc815 \ube44\uc728\uc744 \uad50\ub780\ud569\ub2c8\ub2e4. \ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\uc5d0\ub294 \ud6a8\uacfc\uc801\uc774\uc9c0\ub9cc, \uc774 \ubc29\ubc95\uc740 \uc6d0\ub798 \ud504\ub86c\ud504\ud2b8\ub97c \ucda9\ubd84\ud788 \ubcc0\uacbd\ud558\uc5ec \uc591\uc131 \ud504\ub86c\ud504\ud2b8\uc758 \uc774\ud574 \ub610\ub294 \uac00\ub3c5\uc131 \ubb38\uc81c\ub97c \uc720\ubc1c\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-2108271017&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"looking-at-alternatives\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"looking-at-alternatives\"><span dir=\"auto\">\ub300\uc548\uc744 \uc0b4\ud3b4\ubcf4\ub2e4<\/span><\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">\ubd80\uc988 \uc568\ub7f0\uc740 \uc801\ub300\uc801 \uc811\ubbf8\uc0ac\ub97c \ud3ec\ud568\ud55c \uc77c\ubd80 \ud0c8\uc625 \uacf5\uaca9\uc744 \ucc28\ub2e8\ud558\uba74\uc11c \uc6d0\ub798 \ud504\ub86c\ud504\ud2b8\uc758 \uc758\ubbf8\ub97c \ubcf4\uc874\ud558\ub294 \uc0c8\ub85c\uc6b4 \uc811\uadfc\ubc95\uc744 \ubaa8\uc0c9\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \ud504\ub86c\ud504\ud2b8\uc5d0\uc11c \uc758\ubbf8\uac00 \ub35c\ud55c \ubb38\uc790(\uad6c\ub450\uc810 \ud3ec\ud568, \uacf5\ubc31 \uc81c\uc678)\ub97c \uc81c\uac70\ud569\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\uc774 \ubc29\ubc95\uc774 \ud6a8\uacfc\uc801\uc778 \uc774\uc720\ub294 \ubb34\uc5c7\uc77c\uae4c\uc694? \ud0c8\uc625 \ud504\ub86c\ud504\ud2b8\uc5d0\ub294 \ud0a4\ub9b4 \ubb38\uc790, \uc774\ubaa8\ud2f0\ucf58, \ubcf4\uc774\uc9c0 \uc54a\ub294 \ubb38\uc790, ASCII \ucf54\ub4dc, \ucf54\ub4dc \uad6c\ubb38, \uadf8\ub9ac\uace0 \uae30\ud0c0 \ud754\ud558\uc9c0 \uc54a\uc740 \ubb38\uc790\uac00 \ud3ec\ud568\ub418\ub294 \uacbd\uc6b0\uac00 \ub9ce\ub2e4\ub294 \uc810\uc744 \uace0\ub824\ud558\uc138\uc694. \uc774\ub7ec\ud55c \ubb38\uc790\ub4e4\uc744 \uc81c\uac70\ud568\uc73c\ub85c\uc368 \uc774 \ubc29\ubc95\uc740 \uacf5\uaca9\uc758 \ud6a8\uacfc\ub97c \uac10\uc18c\uc2dc\ud0b5\ub2c8\ub2e4. \ud2b9\ud788, \ubb34\uc791\uc704 \ubb38\uc790\uc640 \ubb38\uc790\uc5f4\ub85c \uad6c\uc131\ub41c GCG \uc811\ubbf8\uc0ac\ub294 \uc774\ub7ec\ud55c \ubb38\uc790\ub4e4\uc744 \uc81c\uac70\ud558\uba74 \ud6a8\uacfc\uac00 \ub5a8\uc5b4\uc9d1\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">SmoothLLM\uacfc \ub2ec\ub9ac, \uc774 \ubc29\uc2dd\uc740 \uc6d0\ub798 \ud504\ub86c\ud504\ud2b8\ub97c \ub300\ubd80\ubd84 \uadf8\ub300\ub85c \uc720\uc9c0\ud558\uace0 \uad6c\ub450\uc810\ub9cc \uc81c\uac70\ud558\ubbc0\ub85c LLM\uc758 \uc774\ud574 \ubc0f \uc801\uc808\ud55c \uc751\ub2f5 \ub2a5\ub825\uc744 \uc800\ud574\ud558\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uae30\ub2a5\uc740 LLM\uc744 \ubbf8\uc158 \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc5d0 \uc0ac\uc6a9\ud558\ub294 \ub370 \ucc28\uc9c8\uc774 \ubc1c\uc0dd\ud558\ub294 \uac83\uc744 \ubc29\uc9c0\ud558\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub429\ub2c8\ub2e4.<\/span><\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1161459501&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"how-jailbreaks-are-changing\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"how-jailbreaks-are-changing\"><span dir=\"auto\">\ud0c8\uc625\uc774 \uc5b4\ub5bb\uac8c \ubcc0\ud654\ud558\uace0 \uc788\ub294\uac00<\/span><\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">\ub2e8\uae30\uc801\uc73c\ub85c LLM\uc5d0 \ub300\ud55c \ud0c8\uc625 \uacf5\uaca9\uc740 \ub354\uc6b1 \uc815\uad50\ud574\uc9c8 \uac83\uc774\uba70, \uacf5\uaca9\uc790\ub294 \ud1a0\ud070 \uc870\uc791 \ubc0f \uc801\ub300\uc801 \ud504\ub86c\ud504\ud2b8 \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1\uc744 \ud1b5\ud574 \ubcf4\uc548 \uba54\ucee4\ub2c8\uc998\uc744 \uc6b0\ud68c\ud560 \uac00\ub2a5\uc131\uc774 \uc788\uc2b5\ub2c8\ub2e4. \uc801\ub300\uc801 \uc785\ub825\uc744 \uc0dd\uc131\ud558\ub294 \uc790\ub3d9\ud654\ub41c \ub3c4\uad6c\uac00 \uae09\uc99d\ud558\uc5ec \uacbd\ud5d8\uc774 \ubd80\uc871\ud55c \uacf5\uaca9\uc790\ub3c4 \ud6a8\uacfc\uc801\uc778 \ud0c8\uc625\uc744 \uad6c\ucd95\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub610\ud55c LLM \uc544\ud0a4\ud14d\ucc98 \ub610\ub294 \ub370\uc774\ud130 \uc138\ud2b8 \ub0b4\uc758 \ud2b9\uc815 \ucde8\uc57d\uc810\uc744 \uc545\uc6a9\ud558\ub294 \ud45c\uc801 \uacf5\uaca9 \ub610\ud55c \uc99d\uac00\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\uc7a5\uae30\uc801\uc73c\ub85c\ub294 \uba38\uc2e0\ub7ec\ub2dd\uc744 \ud65c\uc6a9\ud558\uc5ec \uc2e4\uc2dc\uac04\uc73c\ub85c \uacf5\uaca9\uc744 \uc815\uad50\ud654\ud558\ub294 \uacfc\uc815\uc5d0\uc11c \ub9c8\uc8fc\uce58\ub294 \ubc29\uc5b4 \uccb4\uacc4\uc5d0 \ub300\uc751\ud558\uc5ec \uc801\uc751\ud615 \uacf5\uaca9\uc774 \ub354\uc6b1 \uc99d\uac00\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub525\ud398\uc774\ud06c \ubc0f \uc18c\uc15c \uc5d4\uc9c0\ub2c8\uc5b4\ub9c1\uacfc \uac19\uc740 \uae30\uc220\uc774 \ubc1c\uc804\ud568\uc5d0 \ub530\ub77c, \uc131\uacf5\uc801\uc778 \ud0c8\uc625 \uc720\ub3c4\ub294 \ub354\uc6b1 \uc720\ud574\ud55c \uacb0\uacfc\ub97c \ucd08\ub798\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub610\ud55c, LLM\uc774 \uc911\uc694 \uc2dc\uc2a4\ud15c\uc5d0 \ud544\uc218\uc801\uc774 \ub428\uc5d0 \ub530\ub77c \uaddc\uc81c \ud658\uacbd\uc774 \ubcc0\ud654\ud558\uc5ec \uc774\ub7ec\ud55c \ubcc0\ud654\uc5d0 \uae30\ubc18\ud55c \uc0c8\ub85c\uc6b4 \ud615\ud0dc\uc758 \uacf5\uaca9\uc774 \ubc1c\uc0dd\ud560 \uac00\ub2a5\uc131\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-289122518&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"integrating-defenses-defeating-attackers\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"integrating-defenses-defeating-attackers\"><span dir=\"auto\">\ubc29\uc5b4\ub825 \ud1b5\ud569, \uacf5\uaca9\uc790 \ubb3c\ub9ac\uce58\uae30<\/span><\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span dir=\"auto\">\ud0c8\uc625 \uacf5\uaca9\uc758 \uc9c4\ud654\ub294 LLM \ubc0f \uc9c0\uc6d0 \uc2dc\uc2a4\ud15c\uc758 \ubb34\uacb0\uc131\uc744 \ubcf4\ud638\ud558\uae30 \uc704\ud55c \ud601\uc2e0\uc801\uc778 \ubcf4\uc548 \uc870\uce58\uc758 \ud544\uc694\uc131\uc744 \uac15\uc870\ud569\ub2c8\ub2e4. \uac1c\ubcc4\uc801\uc778 \ub3c4\uad6c\ub9cc\uc73c\ub85c \ud2b9\uc815 \ucde8\uc57d\uc810\uc744 \uc644\ud654\ud560 \uc218\ub294 \uc788\uc9c0\ub9cc, \ub2e4\uba74\uc801\uc778 \uc704\ud611\uc744 \ubc18\ub4dc\uc2dc \ud574\uacb0\ud558\ub294 \uac83\uc740 \uc544\ub2d9\ub2c8\ub2e4. \ub530\ub77c\uc11c \uae30\uad00\uc740 \ud3ec\uad04\uc801\uc774\uace0 \ud1b5\ud569\ub41c \ubcf4\uc548 \uc804\ub7b5\uc744 \ucc44\ud0dd\ud574\uc57c \ud569\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\uc608\ub97c \ub4e4\uc5b4, \ub2e4\uc911 \uc778\uc99d(MFA)\uacfc \uac19\uc740 \uc811\uadfc \uc81c\uc5b4\ub97c \uad6c\ud604\ud558\uba74 \uad8c\ud55c\uc774 \uc788\ub294 \uc9c1\uc6d0\ub9cc LLM\uacfc \uc0c1\ud638 \uc791\uc6a9\ud558\ub3c4\ub85d \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc554\ud638\ud654 \uae30\uc220\uc740 \uc561\uc138\uc2a4 \ud1a0\ud070\uacfc \uc790\uaca9 \uc99d\uba85\uc744 \ub354\uc6b1 \uc548\uc804\ud558\uac8c \ubcf4\ud638\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uba38\uc2e0 \ub7ec\ub2dd\uc744 \uc0ac\uc6a9\ud558\ub294 \uc774\uc0c1 \ud0d0\uc9c0 \uc2dc\uc2a4\ud15c\uc740 LLM \uc0c1\ud638 \uc791\uc6a9\uc744 \ubaa8\ub2c8\ud130\ub9c1\ud558\uace0 \ud0c8\uc625 \ud328\ud134\uc744 \uc2e4\uc2dc\uac04\uc73c\ub85c \uc2dd\ubcc4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub610\ud55c \ucc28\ub4f1 \uac1c\uc778\uc815\ubcf4 \ubcf4\ud638 \uae30\uc220\uc740 \ub370\uc774\ud130 \ucd9c\ub825\uc5d0 \ud1b5\uc81c\ub41c \ub178\uc774\uc988\ub97c \ucd94\uac00\ud558\uc5ec \uc751\ub2f5\uc5d0 \uc601\ud5a5\uc744 \uc8fc\uc9c0 \uc54a\uace0 \ubbfc\uac10\ud55c \uc815\ubcf4\ub97c \ubcf4\ud638\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<p><span dir=\"auto\">\uacc4\uce35\uc801 \uc811\uadfc \ubc29\uc2dd\uc740 \ud558\ub098\uc758 \ubc29\uc5b4 \uba54\ucee4\ub2c8\uc998\uc774 \uc2e4\ud328\ud558\ub354\ub77c\ub3c4 \ub2e4\ub978 \ubc29\uc5b4 \uba54\ucee4\ub2c8\uc998\uc774 \uc2dc\uc2a4\ud15c\uc744 \ubcf4\ud638\ud558\uae30 \uc704\ud574 \uc720\uc9c0\ub418\ub3c4\ub85d \ubcf4\uc7a5\ud569\ub2c8\ub2e4. \uac1c\ubcc4\uc801\uc778 \ubc29\uc5b4 \uccb4\uacc4\ub97c \ub118\uc5b4 \ud1b5\ud569 \uc804\ub7b5\uc73c\ub85c \uc804\ud658\ud558\uba74 \uae30\uad00\ub4e4\uc774 \uac15\ub825\ud55c \ud0c8\uc625 \uacf5\uaca9\uc744 \ud3ec\ud568\ud55c \ubcc0\ud654\ud558\ub294 \uc0ac\uc774\ubc84 \uc704\ud611\uc73c\ub85c\ubd80\ud130 LLM\uc744 \ub354\uc6b1 \ud6a8\uacfc\uc801\uc73c\ub85c \ubcf4\ud638\ud558\ub294 \ub3d9\uc2dc\uc5d0, \uc774\ub7ec\ud55c \ubaa8\ub378\uc744 \uc911\uc694\ud55c \uc784\ubb34\uc758 \uc77c\ubd80\ub85c \uacc4\uc18d \ud65c\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/span><\/p>\n<div id=\"content\" data-events=\"scroll\" data-magellan=\"rz7y9r-magellan\" data-resize=\"content\" data-scroll=\"content\">\n<div class=\"blog-hero-wrapper\" data-aa-hero=\"blog\">\n<div class=\"row\">\n<div class=\"columns large-9 medium-12 small-12\">\n<div class=\"hero-container blog-hero\">\n<div class=\"hero-data blog-hero\">\n<h1>How to Protect LLMs from Jailbreaking Attacks<\/h1>\n<div data-aa-comp-name=\"blog-hero:author-link\">\n<div class=\"cmp-text \">\n<p>By Noah Fleischmann,&nbsp;Amy Wagoner, Ph.D.,&nbsp;and Andre Nguyen, Ph.D.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div class=\"blog-top-container\">\n<div class=\"generic-block\">\n<div class=\"row\">\n<div class=\"small-12 columns\">\n<div class=\"image parbase\">\n<div class=\"image-container \n\n        no-pointer\n        default-style \n        default-style \n        default-style\n\n        \"><img decoding=\"async\" class=\"image widecq-dd-image\" src=\"https:\/\/www.stechstar.com\/user\/zbxe\/files\/attach\/images\/112899\/915\/112\/70eedf6e23c9b9e5c209b083506e56dc.jpg\" alt=\"\" data-autoattach=\"success\" data-emptytext=\"Image\" data-cke-saved-src=\".\/files\/attach\/images\/112899\/915\/112\/70eedf6e23c9b9e5c209b083506e56dc.jpg\"><\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div class=\"sticky-sidebar-container full\">\n<div id=\"sticky-anchor\">\n<div class=\"row\">\n<div class=\"blog-breakout-parsys small-12 large-10 large-offset-1 columns\">\n<div>&nbsp;<\/div>\n<\/div>\n<\/div>\n<div class=\"row\">\n<div class=\"grid-column--center-align columns large-10 full\">\n<div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1628759649&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h2 id=\"booz-allen-examines-how-to-strengthen-model-integrity\" class=\"title title-component title-component--default align-left\">Booz Allen examines how to strengthen model integrity<\/h2>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p><span class=\"rte-large\">Federal agencies are increasingly integrating large language models (LLMs) like Llama-2 and ChatGPT into their operations to streamline tasks and answer questions. Engineers design these models to be \u201chelpful and harmless\u201d and refuse dangerous requests. Techniques like fine-tuning, reinforcement learning with human feedback, and direct preference optimization can further enhance model safety. But despite these measures, a critical LLM vulnerability continues to put AI systems at risk: jailbreak prompts.<\/span><\/p>\n<p>Jailbreak prompts are specific inputs designed to trick LLMs into doing things they shouldn\u2019t. These cleverly designed but malicious prompts can bypass even the most robust security measures, posing significant risks to federal operations.<\/p>\n<p>To help address this challenge, Booz Allen is exploring new defenses against jailbreaking. These approaches can provide agencies with a significant mission advantage: they protect the LLM without hindering its ability to respond to benign prompts so that it can continue functioning as a driver of increased enterprise productivity.<\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1403034696&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"how-jailbreaking-works\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"how-jailbreaking-works\">How Jailbreaking Works<\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p>Adversaries craft jailbreak prompts to manipulate LLMs into revealing sensitive information, such as personally identifiable information, or generating harmful content, such as instructions for illegal activities (e.g., building a bomb, writing a phishing e-mail) or hate speech. Attackers employ a variety of creative techniques, including:<\/p>\n<ul>\n<li>Role Play:&nbsp;Asking the AI to play a specific role or persona to bypass existing guardrails for divulging sensitive or protected information.<\/li>\n<li>Attention Shifting:&nbsp;Distracting the AI to bypass its safety protocols.<\/li>\n<li>Privilege Escalation:&nbsp;Gaining higher-level access within the AI\u2019s system.<\/li>\n<li>Prefix Injection:&nbsp;Adding specific phrases to manipulate the AI\u2019s responses.<\/li>\n<li>Refusal Suppression:&nbsp;Tricking the AI into ignoring its refusal protocols.<\/li>\n<li>Word Games and Obfuscation:&nbsp;Using complex language to confuse the AI.<\/li>\n<li>Multilingual Input and Ciphers:&nbsp;Using different languages or coded messages to bypass security.<\/li>\n<\/ul>\n<p>How effective are these techniques? Research shows&nbsp;<a title=\"Jailbreak Prompts\" href=\"https:\/\/arxiv.org\/abs\/2402.05668\" target=\"_blank\" rel=\"noopener\" data-cke-saved-href=\"https:\/\/arxiv.org\/abs\/2402.05668\">jailbreak prompts consistently achieve high attack success rates<\/a>&nbsp;across various LLMs, including Vicuna, ChatGLM3, GPT-3.5, and PaLM2, underscoring their robustness and transferability despite safety measures. In addition, the&nbsp;<a title=\"automated trees\" href=\"https:\/\/arxiv.org\/abs\/2312.02119\" target=\"_blank\" rel=\"noopener\" data-cke-saved-href=\"https:\/\/arxiv.org\/abs\/2312.02119\">automated tree of attacks with pruning<\/a>&nbsp;(TAP) technique can provide prompts that successfully jailbreak mature LLMs for over 80% of prompts with just a few queries.<\/p>\n<p>Why does jailbreaking work so well? It has to do with conflicting goals within LLMs. On the one hand, the models aim to be helpful by following instructions, but on the other, they need to adhere to safety guidelines to avoid causing harm. In other cases, it may be due to mismatched generalization. That is, the quantity of safety training data cannot fully represent the body of training data, which causes some attack scenarios to be missed. Today, we see a cycle where new jailbreak attacks emerge, prompting the development of new defenses in a never-ending game of cat and mouse.<\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-78211345&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"risk-scenarios-for-government\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"risk-scenarios-for-government\">Risk Scenarios for Government<\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p>Jailbreaking attacks on LLMs pose significant risks to federal agencies. Risks with relevance for national security include data breaches, privacy violations, spread of misinformation, manipulation of automated systems, and compromised decision-making processes.<\/p>\n<p>Imagine a national emergency where a jailbroken LLM responds to prompts with false information about evacuation procedures, causing chaos and endangering lives. Or an attacker uses a jailbroken LLM to trigger a power outage in a major city by injecting malicious commands into a supervisory control and data acquisition system.<\/p>\n<p>Jailbroken LLMs may also bring ethical and legal challenges. Compromised models could generate illegal content, exposing agencies to regulatory penalties. And if a jailbroken LLM produces discriminatory or offensive content that finds its way into an official report, the agency could suffer reputational damage and face lawsuits.&nbsp;<\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1350676556&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"attack-types\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"attack-types\">Attack Types<\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p>Recent advances have enabled more powerful jailbreak prompts by modifying parts of the input text. Security practitioners call this process \u201cperturbation.\u201d Outside of generative AI, researchers have studied how these changes affect text classification problems, like sentiment analysis and toxicity detection. Since detecting harmful inputs and outputs is crucial, it\u2019s important to consider adversarial classification attacks used alongside generative jailbreak prompts. Examples of perturbation-based attacks include:<\/p>\n<ul>\n<li>Character-level perturbations (CLPs):&nbsp;Attackers randomly swap, insert, or delete prompt characters, targeting important words to change their mapping. This removes information the model uses to detect harmful prompts. CLPs may be mistaken for typographical errors or may be virtually invisible (e.g., homoglyphs).<\/li>\n<\/ul>\n<ul>\n<li>Word-level perturbations (WLPs):&nbsp;Attackers substitute important words with synonyms or out-of-vocabulary words to significantly change the model\u2019s behavior.<\/li>\n<\/ul>\n<ul>\n<li>Sentence-level perturbations (SLPs):&nbsp;Attackers rephrase or change the prompt to keep its original meaning but confuse the model. Adding irrelevant sentences or using \u201croundtrip\u201d translations from one language to another and back are common methods.<\/li>\n<\/ul>\n<ul>\n<li>Greedy coordinate gradient (GCG) attacks:&nbsp;GCG attacks maximize the chances that a harmful prompt will produce a targeted affirmative response through the creation of suffixes\u2014strings of seemingly random characters and words\u2014that help bypass the model\u2019s safety features.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-6844399&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"defenses-against-jailbreaking\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"defenses-against-jailbreaking\">Defenses Against Jailbreaking<\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p>Evaluating jailbreaks is very difficult, and mistakes (false positives and false negatives) will happen with any system. While alternative methods, such as using auxiliary models, may decrease the false positive and false negative rate compared to string matching, questions still arise about categorizing samples where the model does not understand the prompt or returns a partial refusal\/compliance. Even the terms \u201cpartial refusal\u201d and \u201cpartial compliance\u201d can be controversial because researchers might disagree on when an output becomes harmful.<\/p>\n<p>Developing a strong defense against jailbreak attacks has also proven challenging due to the variety of models and attack types. Many current defenses target specific types of jailbreaks and often use similar perturbation strategies as the attacks. Some defenses use WLPs and SLPs to reduce the impact of jailbreak prompts. Another approach uses CLPs to minimize the effect of GCG suffixes by randomly changing characters in the adversarial suffix.<\/p>\n<p>However, this method struggles to automatically distinguish between the original prompt and the suffix. The SmoothLLM algorithm, for example, perturbs a percentage of the entire input, including the suffix, rather than isolating the suffix alone. While effective against jailbreak prompts, this approach can alter the original prompt enough to cause comprehension or readability issues for benign prompts.<\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-2108271017&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"looking-at-alternatives\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"looking-at-alternatives\">Looking at Alternatives<\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p>Booz Allen is exploring new approaches that preserve the original prompt\u2019s meaning while thwarting some jailbreak attacks, including adversarial suffixes. They do this by removing less meaningful characters (including punctuation, but not spaces) from the prompt.<\/p>\n<p>Why is this effective? Consider that jailbreak prompts often include Cyrillic characters, emojis, invisible characters, ASCII art, code syntax, and other uncommon characters. By eliminating these, this approach reduces the effectiveness of the attack. Specifically, GCG suffixes, which consist of random characters and strings, become less effective when these characters are removed.<\/p>\n<p>Unlike SmoothLLM, this approach leaves the original prompt mostly intact, removing only punctuation, which does not hinder the LLM\u2019s ability to comprehend and respond appropriately. This capability helps prevent disruptions in the use of LLMs for mission applications.<\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-1161459501&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"how-jailbreaks-are-changing\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"how-jailbreaks-are-changing\">How Jailbreaks Are Changing<\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p>In the near term, jailbreaking attacks on LLMs will become more sophisticated, with adversaries potentially bypassing safety mechanisms through token manipulation and adversarial prompt engineering. Automated tools for generating adversarial inputs may proliferate, allowing inexperienced attackers to craft effective jailbreaks. And targeted attacks exploiting specific vulnerabilities within LLM architectures or datasets may also increase.<\/p>\n<p>Over the long term, adaptive attacks could further increase in response to encountered defenses, with machine learning used to refine attacks in real time. With improving technologies like deepfakes and social engineering, successful jailbreak prompts might lead to more harmful outputs. In addition, as LLMs become integral to critical systems, the regulatory landscape could shift, potentially giving rise to new forms of attacks based on these changes.<\/p>\n<\/div>\n<\/div>\n<div class=\"title\">\n<div class=\"title-component-wrapper\" data-cmp-data-layer=\"{\n  &quot;comp-289122518&quot; : {\n    &quot;title&quot; : null,\n    &quot;url&quot; : &quot;&quot;,\n    &quot;type&quot; : &quot;boozallen\/components\/general\/title&quot;\n  }\n}\"><\/p>\n<h3 id=\"integrating-defenses-defeating-attackers\" class=\"title title-component title-component--default align-left in-page-nav-target\" data-magellan-target=\"integrating-defenses-defeating-attackers\">Integrating Defenses, Defeating Attackers<\/h3>\n<\/div>\n<\/div>\n<div class=\"text\">\n<div class=\"cmp-text \">\n<p>The evolution in jailbreak attacks underscores the need for innovative security measures to safeguard the integrity of LLMs and supported systems. While isolated tools can mitigate specific vulnerabilities, they don\u2019t necessarily address multifaceted threats. As a result, agencies should adopt a comprehensive, integrated security strategy.<\/p>\n<p>For example, implementing access controls like multifactor authentication can help ensure that only authorized personnel interact with LLMs. Cryptographic techniques can further secure access tokens and credentials. Anomaly detection systems that use machine learning can monitor LLM interactions and identify jailbreak patterns in real time. And differential privacy techniques can add controlled noise to data outputs, protecting sensitive information without affecting responses.<\/p>\n<p>A layered approach ensures that, even if one defense mechanism fails, others remain in place to protect the system. Moving beyond isolated defenses to an integrated strategy can help agencies more effectively safeguard LLMs against shifting cyber threats, including potent jailbreaking attacks, while continuing to harness these models as part of critical missions.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>[\ucd9c\ucc98]&nbsp;https:\/\/www.boozallen.com\/insights\/ai-research\/how-to-protect-llms-from-jailbreaking-attacks.html<\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_11417\" class=\"pvc_stats all  \" data-element-id=\"11417\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg aria-hidden=\"true\" focusable=\"false\" data-prefix=\"far\" data-icon=\"chart-bar\" role=\"img\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" viewBox=\"0 0 512 512\" class=\"svg-inline--fa fa-chart-bar fa-w-16 fa-2x\"><path fill=\"currentColor\" d=\"M396.8 352h22.4c6.4 0 12.8-6.4 12.8-12.8V108.8c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v230.4c0 6.4 6.4 12.8 12.8 12.8zm-192 0h22.4c6.4 0 12.8-6.4 12.8-12.8V140.8c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v198.4c0 6.4 6.4 12.8 12.8 12.8zm96 0h22.4c6.4 0 12.8-6.4 12.8-12.8V204.8c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v134.4c0 6.4 6.4 12.8 12.8 12.8zM496 400H48V80c0-8.84-7.16-16-16-16H16C7.16 64 0 71.16 0 80v336c0 17.67 14.33 32 32 32h464c8.84 0 16-7.16 16-16v-16c0-8.84-7.16-16-16-16zm-387.2-48h22.4c6.4 0 12.8-6.4 12.8-12.8v-70.4c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v70.4c0 6.4 6.4 12.8 12.8 12.8z\" class=\"\"><\/path><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/www.stechstar.com\/user\/wordpress\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>[\uc778\uacf5\uc9c0\ub2a5 \uc11c\ube44\uc2a4 \ubcf4\uc548] \ud0c8\uc625 \uacf5\uaca9\uc73c\ub85c\ubd80\ud130 LLM\uc744 \ubcf4\ud638\ud558\ub294 \ubc29\ubc95 , How to Protect LLMs from Jailbreaking Attacks \ud0c8\uc625 \uacf5\uaca9\uc73c\ub85c\ubd80\ud130 LLM\uc744 \ubcf4\ud638\ud558\ub294 \ubc29\ubc95 \ub178\uc544 \ud50c\ub77c\uc774\uc288\ub9cc, \uc5d0\uc774\ubbf8 \uc640\uace0\ub108 \ubc15\uc0ac, \uc559\ub4dc\ub808 \uc751\uc6b0\uc60c \ubc15\uc0ac &nbsp; Booz Allen\uc740 \ubaa8\ub378 \ubb34\uacb0\uc131\uc744 \uac15\ud654\ud558\ub294 \ubc29\ubc95\uc744 \uc870\uc0ac\ud569\ub2c8\ub2e4. \uc5f0\ubc29 \uae30\uad00\ub4e4\uc740 \uc5c5\ubb34\ub97c \uac04\uc18c\ud654\ud558\uace0 \uc9c8\ubb38\uc5d0 \ub2f5\ud558\uae30 \uc704\ud574 Llama-2 \ubc0f ChatGPT\uc640 \uac19\uc740 \ub300\uaddc\ubaa8 \uc5b8\uc5b4 \ubaa8\ub378(LLM)\uc744 \uc6b4\uc601\uc5d0 \uc810\uc810 \ub354 \ud1b5\ud569\ud558\uace0 [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_11417\" class=\"pvc_stats all  \" data-element-id=\"11417\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg aria-hidden=\"true\" focusable=\"false\" data-prefix=\"far\" data-icon=\"chart-bar\" role=\"img\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" viewBox=\"0 0 512 512\" class=\"svg-inline--fa fa-chart-bar fa-w-16 fa-2x\"><path fill=\"currentColor\" d=\"M396.8 352h22.4c6.4 0 12.8-6.4 12.8-12.8V108.8c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v230.4c0 6.4 6.4 12.8 12.8 12.8zm-192 0h22.4c6.4 0 12.8-6.4 12.8-12.8V140.8c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v198.4c0 6.4 6.4 12.8 12.8 12.8zm96 0h22.4c6.4 0 12.8-6.4 12.8-12.8V204.8c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v134.4c0 6.4 6.4 12.8 12.8 12.8zM496 400H48V80c0-8.84-7.16-16-16-16H16C7.16 64 0 71.16 0 80v336c0 17.67 14.33 32 32 32h464c8.84 0 16-7.16 16-16v-16c0-8.84-7.16-16-16-16zm-387.2-48h22.4c6.4 0 12.8-6.4 12.8-12.8v-70.4c0-6.4-6.4-12.8-12.8-12.8h-22.4c-6.4 0-12.8 6.4-12.8 12.8v70.4c0 6.4 6.4 12.8 12.8 12.8z\" class=\"\"><\/path><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/www.stechstar.com\/user\/wordpress\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ngg_post_thumbnail":0,"footnotes":""},"categories":[15,25,24],"tags":[],"class_list":["post-11417","post","type-post","status-publish","format-standard","hentry","category-stechstar-com-","category-25","category-24"],"a3_pvc":{"activated":true,"total_views":14,"today_views":0},"_links":{"self":[{"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/posts\/11417","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/comments?post=11417"}],"version-history":[{"count":3,"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/posts\/11417\/revisions"}],"predecessor-version":[{"id":11863,"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/posts\/11417\/revisions\/11863"}],"wp:attachment":[{"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/media?parent=11417"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/categories?post=11417"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.stechstar.com\/user\/wordpress\/wp-json\/wp\/v2\/tags?post=11417"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}