[{"_1":2,"_129":-5,"_130":-5},"loaderData",{"_3":4,"_9":10},"root",{"_5":6,"_7":8},"locale","id","isCms",false,"routes/tags/$slug",{"_11":12,"_13":14,"_42":43,"_117":118,"_123":120,"_127":39,"_128":-5},"slug","anthropic","ads",[15],{"_6":16,"_17":18,"_19":20,"_21":22,"_23":24,"_25":26,"_27":28,"_29":30,"_28":31,"_32":33,"_34":35,"_36":37,"_38":39,"_40":41},"019cc7ba-cc99-73c7-8531-f093038fe91a","position_id","2f16d5c6-a732-4e00-b8c6-6306c0d32904","position_name","Sidebar Middle","position_code","sidebar_middle","name","Membaca 1","code","Membaca 2","type","image","link","https://indonesiawindow.com/lighterasi","banners/LNl7ZGDdnD0VTU52UMOugAjr6ns9lzdIRN89W13M.png","is_active",true,"description","","image_url","https://indonesiawindow.sgp1.cdn.digitaloceanspaces.com/banners/LNl7ZGDdnD0VTU52UMOugAjr6ns9lzdIRN89W13M.png","visitor",12,"created_at","2026-03-07","posts",[44],{"_6":45,"_46":47,"_11":48,"_49":50,"_51":52,"_53":54,"_55":56,"_57":58,"_59":60,"_61":62,"_63":64,"_65":66,"_67":68,"_69":70,"_71":35,"_72":35,"_73":-5,"_28":74,"_75":35,"_76":77,"_36":78,"_79":80,"_81":82,"_114":115,"_40":116},"019fba5b-76c8-722c-9c0b-d7ed10004eef","title","Gara-gara internet tak diputus, AI Anthropic 'menyusup' ke sistem tiga organisasi nyata saat uji keamanan siber","gara-gara-internet-tak-diputus-ai-anthropic-menyusup-ke-sistem-tiga-organisasi-nyata-saat-uji-keamanan-siber","main_id","019bb73f-297b-7103-96b7-87ea173dc182","main_name","Iptek","main_slug","iptek-id","date","2026-08-01 00:00:00","author_id","019bb745-b426-73a2-9a49-5ddcb6765ea0","author_name","indowin","full_name","Indonesia Window","author_firstname","Indonesia","author_slug","indonesiawindow","content","

San Francisco, Amerika Serikat (Xinhua/Indonesia Window) – Perusahaan kecerdasan buatan (artificial intelligence/AI) Amerika Serikat (AS), Anthropic, pada Kamis (30/7) mengatakan tiga model AI-nya mengakses atau berinteraksi dengan sistem komputer milik tiga organisasi nyata selama evaluasi kemampuan siber internal, setelah akses internet secara tidak sengaja dibiarkan tersedia.

Pengungkapan tersebut disampaikan dalam sebuah unggahan blog perusahaan berjudul \"Menyelidiki tiga insiden di dunia nyata dalam evaluasi keamanan siber kami\" (Investigating three real-world incidents in our cybersecurity evaluations).

Menurut Anthropic, insiden-insiden tersebut teridentifikasi setelah perusahaan itu meninjau lebih dari 141.000 transkrip evaluasi kemampuan siber menyusul pengungkapan insiden keamanan serupa oleh OpenAI.

Perusahaan tersebut menyatakan model-model itu telah diberi instruksi bahwa mereka beroperasi di lingkungan tersimulasi tanpa akses internet.

Namun, karena adanya kesalahpahaman antara Anthropic dan mitra evaluasinya, Irregular, akses internet secara tidak sengaja dibiarkan tersedia, sehingga memungkinkan model-model itu berinteraksi dengan sistem eksternal yang secara keliru mereka anggap sebagai bagian dari lingkungan evaluasi.

Ketiga insiden itu melibatkan Claude Opus 4.7, Claude Mythos 5, dan sebuah model uji penelitian internal dalam latihan capture-the-flag (CTF) yang dirancang untuk menilai kemampuan siber ofensif, menurut perusahaan tersebut.

Anthropic menyatakan pihaknya segera memberi tahu organisasi-organisasi yang terdampak setelah mengonfirmasi insiden tersebut. Dalam dua dari tiga kasus, organisasi-organisasi tersebut tidak menyadari bahwa sistem mereka telah diakses sebelum menerima pemberitahuan dari Anthropic.

Menurut perusahaan itu, model-model tersebut mengeksploitasi kelemahan keamanan umum, termasuk kata sandi yang lemah dan layanan yang tidak terautentikasi, alih-alih kerentanan perangkat lunak yang sebelumnya tidak diketahui.

Anthropic menyatakan pihaknya tidak menemukan bukti bahwa insiden-insiden tersebut menyebabkan kerugian jangka panjang atau mengakibatkan pencurian atau kebocoran informasi sensitif.

Perusahaan itu menyatakan telah menangguhkan semua evaluasi kemampuan siber yang berpotensi mengakses internet publik sembari menerapkan langkah-langkah pengamanan tambahan dan melakukan peninjauan yang lebih luas terhadap prosedur evaluasinya.

Anthropic menyatakan bertanggung jawab penuh atas insiden-insiden itu dan akan menggunakan temuan-temuan ini guna semakin memperkuat langkah-langkah pengendalian untuk evaluasi keselamatan AI di masa mendatang.

Laporan: Redaksi

","excerpt","San Francisco, Amerika Serikat (Xinhua/Indonesia Window) – Perusahaan kecerdasan buatan (artificial intelligence/AI) Amerika Serikat (AS), A...","seo_kw","seo_title","seo_description","articles/KDUfGHd1WFeoN8zJtxW991ZU5WB9RRXgbjPAshzA.jpg","image_alt","image_caption","Orang-orang menyaksikan robot humanoid bermain sepak bola di KOID Shop di New York City, Amerika Serikat, pada 25 Juni 2026. (Xinhua/Vivien Liu)","https://indonesiawindow.sgp1.cdn.digitaloceanspaces.com/articles/KDUfGHd1WFeoN8zJtxW991ZU5WB9RRXgbjPAshzA.jpg","status","publish","tags",[83,93,100,107],{"_6":84,"_85":-5,"_23":86,"_11":12,"_28":-5,"_34":-5,"_40":87,"_88":87,"_89":90},"019fba5b-76ce-7029-9dcb-764567b36a3d","old_id","Anthropic","2026-07-31T22:46:24.000000Z","updated_at","pivot",{"_91":45,"_92":84,"_40":87,"_88":87},"post_id","tag_id",{"_6":94,"_85":95,"_23":96,"_11":97,"_28":-5,"_34":-5,"_40":98,"_88":98,"_89":99},"019bb744-6c70-71f1-8e3b-2880e8626e5f",9907,"artificial intelligence","artificial-intelligence","2026-01-13T12:11:13.000000Z",{"_91":45,"_92":94,"_40":87,"_88":87},{"_6":101,"_85":102,"_23":103,"_11":104,"_28":-5,"_34":-5,"_40":105,"_88":105,"_89":106},"019bb744-ad4b-722f-a206-f3f4fab8ee0b",16922,"keamanan siber","keamanan-siber","2026-01-13T12:11:29.000000Z",{"_91":45,"_92":101,"_40":87,"_88":87},{"_6":108,"_85":109,"_23":110,"_11":111,"_28":-5,"_34":-5,"_40":112,"_88":112,"_89":113},"019bb744-ff28-70f2-9a1f-23d50f89f9e2",26019,"OpenAI","openai","2026-01-13T12:11:50.000000Z",{"_91":45,"_92":108,"_40":87,"_88":87},"comments",[],"2026-07-31","pagination",{"_119":120,"_121":120,"_122":39,"_123":120,"_124":120,"_125":126},"count",1,"total","perPage","currentPage","totalPages","links",{},"pageSize","error","actionData","errors"]