{"id":7156,"date":"2025-07-22T06:37:00","date_gmt":"2025-07-21T21:37:00","guid":{"rendered":"https:\/\/devneko.jp\/wordpress\/?p=7156"},"modified":"2025-07-20T22:41:19","modified_gmt":"2025-07-20T13:41:19","slug":"the-devil-behind-the-mask-an-emergent-safety-vulnerability-of-diffusion-llms","status":"publish","type":"post","link":"https:\/\/devneko.jp\/wordpress\/?p=7156","title":{"rendered":"The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs\u00a0"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li><strong>The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs\u00a0<\/strong>[39.9]<br>DLLM\u306e\u30e6\u30cb\u30fc\u30af\u306a\u5b89\u5168\u6027\u306e\u5f31\u70b9\u3092\u751f\u304b\u3057\u305f\u3001\u6700\u521d\u306e\u7cfb\u7d71\u7684\u306a\u7814\u7a76\u304a\u3088\u3073\u8131\u7344\u653b\u6483\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3067\u3042\u308bDIJA\u3092\u63d0\u6848\u3059\u308b\u3002 \u63d0\u6848\u3059\u308bDIJA\u306f,dLLM\u306e\u30c6\u30ad\u30b9\u30c8\u751f\u6210\u6a5f\u69cb\u3092\u5229\u7528\u3057\u305f\u5bfe\u5411\u7684\u30a4\u30f3\u30bf\u30fc\u30ea\u30fc\u30d6\u30fb\u30de\u30b9\u30af\u30c6\u30ad\u30b9\u30c8\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u69cb\u7bc9\u3059\u308b\u3002 \u672c\u7814\u7a76\u306f, \u65b0\u305f\u306a\u8a00\u8a9e\u30e2\u30c7\u30eb\u306b\u304a\u3044\u3066, \u5b89\u5168\u30a2\u30e9\u30a4\u30e1\u30f3\u30c8\u306e\u518d\u8003\u306e\u5fc5\u8981\u6027\u3092\u6d6e\u304d\u5f6b\u308a\u306b\u3059\u308b\u3082\u306e\u3067\u3042\u308b\u3002<br><a href=\"http:\/\/arxiv.org\/abs\/2507.11097v1\">\u8ad6\u6587<\/a>\u00a0\u00a0<a href=\"https:\/\/fugumt.com\/fugumt\/paper_check\/2507.11097v1\">\u53c2\u8003\u8a33\uff08\u30e1\u30bf\u30c7\u30fc\u30bf\uff09<\/a>\u00a0 \u00a0(Tue, 15 Jul 2025 08:44:46 GMT)<\/li>\n\n\n\n<li>dLLM\u306b\u5bfe\u3059\u308b\u653b\u6483\u624b\u6cd5\u306e\u63d0\u6848\u3002\u300cBy interleaving sets of [MASK] tokens after vanilla malicious prompt, as shown in Figure 2, a dLLM is coerced into generating harmful instructions purely to maintain contextual consistency.  Moreover, in contrast to autoregressive LLMs, which generate tokens sequentially and can perform on-the-fly rejection of unsafe continuations, dLLMs decode masked tokens in parallel at each step, substantially limiting the model\u2019s ability to conduct dynamic risk assessment or intervene during generation (e g , reject sampling for tokens corresponding to harmful contents). Consequently, defenses designed for left-to-right models break down, opening the door to powerful new jailbreak attacks.\u300d\u3068\u3042\u308b\u901a\u308a\u3001CausalLM\u3068\u306f\u5225\u4f53\u7cfb\u3067\u3042\u308b\u30e2\u30c7\u30eb\u306e\u7279\u5fb4\u3092\u5229\u7528\u3057\u305f\u653b\u6483\u624b\u6cd5\u3068\u306a\u3063\u3066\u3044\u3066\u3001\u653b\u6483\u6210\u529f\u7387\u3082\u9ad8\u3044\u3002<\/li>\n\n\n\n<li>\u30ea\u30dd\u30b8\u30c8\u30ea\u306f<a href=\"https:\/\/github.com\/ZichenWen1\/DIJA\">GitHub &#8211; ZichenWen1\/DIJA: code for &#8220;The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs&#8221;<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[32,114,223],"class_list":["post-7156","post","type-post","status-publish","format-standard","hentry","category-arxiv","tag-attack","tag-diffusion-model","tag-llm"],"_links":{"self":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7156","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7156"}],"version-history":[{"count":1,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7156\/revisions"}],"predecessor-version":[{"id":7157,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/7156\/revisions\/7157"}],"wp:attachment":[{"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7156"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7156"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/devneko.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7156"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}