
{"id":13074,"date":"2026-10-09T16:38:06","date_gmt":"2026-10-09T08:38:06","guid":{"rendered":"https:\/\/infernews.com\/blog\/advsim2real-training-web-agents-against-adaptive-prompt-injection-in-a-web-world-model\/"},"modified":"2026-10-09T16:46:00","modified_gmt":"2026-10-09T08:46:00","slug":"advsim2real-training-web-agents-against-adaptive-prompt-injection-in-a-web-world-model","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/advsim2real-training-web-agents-against-adaptive-prompt-injection-in-a-web-world-model\/","title":{"rendered":"AdvSim2Real \u8b93\u7db2\u9801\u722c\u87f2\u4ee3\u7406\u5b78\u61c2\u62b5\u6297\u63d0\u793a\u6ce8\u5165"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/pasted-143a361f5d59.jpg\" alt=\"AdvSim2Real overview\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7576\u7db2\u9801\u5167\u6587\u540c\u6642\u5305\u542b\u4efb\u52d9\u8cc7\u6599\u548c\u60e1\u610f\u6307\u4ee4\uff0c\u8981\u4ee3\u7406\u5b8c\u5168\u5ffd\u7565\u975e\u5e38\u91cd\u8981\u3002AdvSim2Real \u5c6c\u65bc\u7db2\u9801\u4ee3\u7406\u5b89\u5168\u8a13\u7df4\u9805\u76ee\uff0c\u900f\u904e\u52d5\u614b\u7522\u751f\u4efb\u52d9\u548c\u63d0\u793a\u6ce8\u5165\uff0c\u8655\u7406\u4ee3\u7406\u88ab\u9801\u9762\u5167\u5bb9\u5e36\u504f\u7684\u554f\u984c\uff0c\u4e26\u628a\u80fd\u529b\u63d0\u5347\u5e36\u5230\u771f\u5be6 Chromium \u700f\u89bd\u5668\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9805\u76ee\u5728\u51cd\u7d50\u7684 WebWorld-14B \u4e2d\u5171\u540c\u8a13\u7df4\u4efb\u52d9\u8ab2\u7a0b\u3001\u6ce8\u5165\u5c0d\u624b\u548c\u7db2\u9801\u4ee3\u7406\u3002\u8ab2\u7a0b\u5c08\u6311\u4ee3\u7406\u7d04\u6709\u4e00\u534a\u6a5f\u6703\u5b8c\u6210\u7684\u4efb\u52d9\uff0c\u5c0d\u624b\u53ea\u6709\u5728\u6ce8\u5165\u4ee4\u6210\u529f\u8b8a\u6210\u5931\u6557\u6642\u624d\u7372\u5f97\u56de\u5831\uff1b\u6bcf\u689d\u64cd\u4f5c\u8ecc\u8de1\u518d\u7531\u51cd\u7d50\u7684 Qwen3.8-27B \u8a55\u5be9\u3002\u57f7\u884c\u4ee3\u7406\u63a1\u7528 Qwen3.5-4B\uff0c\u8a13\u7df4\u5f8c\u65e2\u63d0\u5347\u6b63\u5e38\u4efb\u52d9\u80fd\u529b\uff0c\u4e5f\u6e1b\u5c11\u5c0d\u653b\u64ca\u7684\u654f\u611f\u5ea6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img data-dominant-color=\"e4e1e1\" data-has-transparency=\"false\" style=\"--dominant-color: #e4e1e1;\" loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"359\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/AdvSim2Real.jpg\" alt=\"\" class=\"wp-image-13080 not-transparent\" srcset=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/AdvSim2Real.jpg 1024w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/AdvSim2Real-300x105.jpg 300w, https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/AdvSim2Real-767x269.jpg 767w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th class=\"has-text-align-left\" data-align=\"left\">\u5143\u4ef6<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u8ad6\u6587\u914d\u7f6e\uff0f\u4f5c\u7528<\/th><\/tr><\/thead><tbody><tr><td class=\"has-text-align-left\" data-align=\"left\">Executor<\/td><td class=\"has-text-align-left\" data-align=\"left\">Qwen3.5-4B\uff1b\u57f7\u884c\u7db2\u9801\u4efb\u52d9<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\">Curriculum<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u7522\u751f\u7b26\u5408 Agent \u7576\u524d\u80fd\u529b\u5605\u8a13\u7df4\u4efb\u52d9<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\">Adversary<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u7522\u751f\u4ee4 Agent \u504f\u96e2\u4f7f\u7528\u8005\u76ee\u6a19\u5605\u7db2\u9801\u6ce8\u5165\u5167\u5bb9<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\">World model<\/td><td class=\"has-text-align-left\" data-align=\"left\">WebWorld-14B\uff1b\u6a21\u64ec\u6bcf\u6b21\u52d5\u4f5c\u5f8c\u5605\u4e0b\u4e00\u500b\u7db2\u9801\u72c0\u614b<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\">Judge<\/td><td class=\"has-text-align-left\" data-align=\"left\">Qwen3.8-27B\uff1b\u5224\u65b7\u4efb\u52d9\u662f\u5426\u5b8c\u6210<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">150 \u9805\u6e2c\u8a66\u4efb\u52d9\u4e2d\uff0c\u6b63\u5e38\u5b8c\u6210\u7387\u7531 74.89%\u5347\u81f3 81.33%\uff1b\u9762\u5c0d\u4e09\u500b\u8a13\u7df4\u6240\u5f97\u5c0d\u624b\u6642\uff0c\u5b8c\u6210\u7387\u7531 48.07%\u5347\u81f3 57.48%\u3002\u5c0d\u672a\u53c3\u8207\u8a13\u7df4\u7684 Kimi-K3 \u5c0d\u624b\uff0c\u5b8c\u6210\u7387\u76f8\u5c0d\u57fa\u790e\u4ee3\u7406\u63d0\u5347 33.6%\uff1b\u5728\u4e0d\u547c\u53eb\u4e16\u754c\u6a21\u578b\u7684\u771f\u5be6 Chromium \u6e2c\u8a66\u4e2d\uff0c\u56b4\u683c\u6210\u529f\u7387\u5247\u7531 25.56%\u5347\u81f3 44.44%\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8a13\u7df4\u65b9\u5f0f\uff1a<\/strong>\u4efb\u52d9\u3001\u653b\u64ca\u548c\u57f7\u884c\u4ee3\u7406\u540c\u6b65\u6f14\u5316<\/li>\n\n\n\n<li><strong>\u6838\u5fc3\u9650\u5236\uff1a<\/strong>\u7d50\u679c\u9ad8\u5ea6\u4f9d\u8cf4 WebWorld-14B \u6a21\u64ec\u8cea\u7d20\u53ca LLM \u8a55\u5be9\u6e96\u78ba\u5ea6<\/li>\n\n\n\n<li><strong>\u6cdb\u5316\u80fd\u529b\uff1a<\/strong>\u80fd\u62b5\u6297\u672a\u53c3\u8207\u8a13\u7df4\u7684 Kimi-K3 \u5c0d\u624b<\/li>\n\n\n\n<li><strong>\u9a57\u8b49\u7bc4\u570d\uff1a<\/strong>\u6db5\u84cb 150 \u9805\u7db2\u9801\u4efb\u52d9\u53ca\u771f\u5be6 Chromium \u6e2c\u8a66<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5132\u5b58\u5eab\u63d0\u4f9b\u7a0b\u5f0f\u78bc\u3001benchmark\u3001\u8a13\u7df4 checkpoint \u548c\u7d50\u679c\uff0c\u8f03\u9069\u5408\u7814\u7a76\u7db2\u9801\u4ee3\u7406\u5b89\u5168\u3001Computer-use agents \u53ca\u5f37\u5316\u5b78\u7fd2\u7684\u5718\u968a\u91cd\u73fe\u3002\u4f7f\u7528\u8005\u4e0d\u80fd\u628a\u5b83\u7576\u6210\u5373\u88dd\u5373\u7528\u7684\u700f\u89bd\u5668\u9632\u8b77\u5de5\u5177\uff1b\u8981\u91cd\u73fe\u8a13\u7df4\u6216\u6bd4\u8f03\u7d50\u679c\uff0c\u4ecd\u9700\u6309\u9805\u76ee\u63d0\u4f9b\u7684\u6a21\u64ec\u5668\u3001\u6a21\u578b checkpoint \u548c\u8a55\u5be9\u6d41\u7a0b\u5efa\u7acb\u6e2c\u8a66\u74b0\u5883\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/sarim-mbzuai.github.io\/advsim2real\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/Sarim-MBZUAI\/advsim2real\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/Sarim-Hash\/advsim2real-stage1-curr1epoch-exec2epochs-iter3-nprop150\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7db2\u9801\u722c\u87f2\u4ee3\u7406\u4e0d\u80fd\u5b8c\u5168\u5ffd\u7565\u9801\u9762\u5167\u5bb9\uff0cAdvSim2Real \u6539\u7528\u81ea\u6211\u5c0d\u6297\u8a13\u7df4\uff0c\u4ee4\u4ee3\u7406\u5728\u4fdd\u7559\u64cd\u4f5c\u80fd\u529b\u4e4b\u9918\u66f4\u80fd\u62b5\u79a6\u60e1\u610f\u6307\u4ee4\u3002<\/p>\n","protected":false},"author":8,"featured_media":13073,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,116,76,186,127,185,178,38,150,170,199,207,214],"tags":[],"class_list":["post-13074","post","type-post","status-publish","format-standard","hentry","category-133","category-agentic","category-76","category-186","category-127","category-qwen","category-google","category-38","category-150","category-170","category-dataset-","category-kimi","category-reinforcement-learning"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13074","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=13074"}],"version-history":[{"count":3,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13074\/revisions"}],"predecessor-version":[{"id":13081,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13074\/revisions\/13081"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13073"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13074"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13074"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13074"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}