{"id":144022,"date":"2026-08-18T19:51:09","date_gmt":"2026-08-18T19:51:09","guid":{"rendered":"https:\/\/www.europesays.com\/ai\/144022\/"},"modified":"2026-08-18T19:51:09","modified_gmt":"2026-08-18T19:51:09","slug":"openai-lays-out-new-security-changes-after-its-ai-hacked-hugging-face","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ai\/144022\/","title":{"rendered":"OpenAI lays out new security changes after its AI hacked Hugging Face"},"content":{"rendered":"<p class=\"duet--article--dangerously-set-cms-markup duet--article--standard-paragraph _1044qizi _18mzr4b1 _18mzr4b0 _19wv7tc1\">OpenAI is announcing <a href=\"https:\/\/openai.com\/index\/pacing-model-development-cyber-capabilities\/\" rel=\"nofollow noopener\" target=\"_blank\">security updates<\/a> following the July news that its AI broke out of a sandboxed environment and <a href=\"https:\/\/www.theverge.com\/ai-artificial-intelligence\/968988\/openai-hugging-face-hack-ai\" rel=\"nofollow noopener\" target=\"_blank\">accidentally hacked Hugging Face<\/a>, including improvements to its research environments, monitoring, and alignment techniques. The company had already put the <a href=\"https:\/\/www.theverge.com\/ai-artificial-intelligence\/976948\/openai-astra-model-pause-critical-cyber-capabilities\" rel=\"nofollow noopener\" target=\"_blank\">brakes on a new model, Astra<\/a>, that it thinks could have \u201ccritical\u201d cybersecurity capabilities, and the company says it instituted a two-week pause in reinforcement learning (RL) training on its \u201clatest models intended for deployment\u201d while it tightened up security. The company\u2019s \u201clargest planned frontier RL run remains on hold.\u201d<\/p>\n<p class=\"duet--article--dangerously-set-cms-markup duet--article--standard-paragraph _1044qizi _18mzr4b1 _18mzr4b0 _19wv7tc1\">For its frontier model research, OpenAI now requires stronger sandboxes for workloads that \u201cexecute model-generated or otherwise untrusted code,\u201d and has more controls to \u201cisolate higher-risk and untrusted workloads from the internet.\u201d It has also updated its research environment to \u201cremove potentially vulnerable shared services, reduce standing privileges, and improve security and trust boundaries.\u201d<\/p>\n<p class=\"duet--article--dangerously-set-cms-markup duet--article--standard-paragraph _1044qizi _18mzr4b1 _18mzr4b0 _19wv7tc1\">As part of the company\u2019s expanded monitoring setup, OpenAI now aims to issue an alert \u201cwithin 30 minutes after concerning activity is surfaced,\u201d OpenAI says. If the people paged after an alert can\u2019t \u201cconclusively\u201d determine whether an alert is a false positive within 30 minutes, \u201cthose teams are expected to pause the activity.\u201d<\/p>\n<p class=\"duet--article--dangerously-set-cms-markup duet--article--standard-paragraph _1044qizi _18mzr4b1 _18mzr4b0 _19wv7tc1\">OpenAI also says that it\u2019s applying \u201cour core alignment techniques across more stages of the training process,\u201d including reward models that \u201cbetter detect and discourage unsafe behavior\u201d and training models \u201cto be more honest about their actions, capabilities, and limitations.\u201d<\/p>\n<p class=\"duet--article--dangerously-set-cms-markup duet--article--standard-paragraph _1044qizi _18mzr4b1 _18mzr4b0 _19wv7tc1\">Since the discovery of the Hugging Face breach, <a href=\"https:\/\/www.theverge.com\/ai-artificial-intelligence\/973670\/anthropic-claude-hacked-organizations-during-cyber-tests\" rel=\"nofollow noopener\" target=\"_blank\">Anthropic<\/a> and <a href=\"https:\/\/www.theverge.com\/ai-artificial-intelligence\/976040\/now-metas-ai-agents-are-going-rogue\" rel=\"nofollow noopener\" target=\"_blank\">Meta<\/a> have also found that their AI models had hacked other organizations.<\/p>\n","protected":false},"excerpt":{"rendered":"OpenAI is announcing security updates following the July news that its AI broke out of a sandboxed environment&hellip;\n","protected":false},"author":2,"featured_media":14821,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[24,66,157,314,781],"class_list":["post-144022","post","type-post","status-publish","format-standard","has-post-thumbnail","category-openai","tag-ai","tag-news","tag-openai","tag-security","tag-tech"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/144022","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=144022"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/144022\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/14821"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=144022"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=144022"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=144022"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}