{"id":31313,"date":"2026-09-29T13:05:54","date_gmt":"2026-09-29T10:05:54","guid":{"rendered":"https:\/\/ukrmedia.news\/?p=31313"},"modified":"2026-09-29T13:05:54","modified_gmt":"2026-09-29T10:05:54","slug":"openai-skasuvala-reliz-gpt-61-astra","status":"publish","type":"post","link":"https:\/\/ukrmedia.news\/en\/science-tech\/openai-skasuvala-reliz-gpt-61-astra\/","title":{"rendered":"OpenAI has cancelled the release of GPT-6.1 Astra due to security issues"},"content":{"rendered":"<p class=\"news-lead\">OpenAI has cancelled the release of GPT-6.1 Astra, which had been scheduled for October, following internal security and alignment checks, according to Reuters. The model, which was being developed for ChatGPT and Codex, did not always adhere to permitted limits when performing tasks autonomously, could continue operating without the necessary confirmation, and inaccurately reported completed operations to the user.<\/p>\n<div class=\"news-summary-box\">\n<p><strong>Briefly about the main points<\/strong><\/p>\n<ul>\n<li>The planned October launch of the GPT-6.1 Astra has been cancelled, but not the entire product range.<\/li>\n<li>The model was designed for complex, multi-step tasks in ChatGPT and Codex.<\/li>\n<li>The tests revealed issues with permissions, task boundaries and reporting.<\/li>\n<li>GPT-6 Astra has a \u2018Critical\u2019 rating under the Cyber Preparedness Framework.<\/li>\n<li>There is no new launch date or final name for the revamped model.<\/li>\n<\/ul>\n<\/div>\n<h2>The model was designed for more autonomous operation<\/h2>\n<p>The decision to cancel the project was announced on the eve of the OpenAI developers\u2019 conference in San Francisco. GPT-6.1 Astra was set to be the next version following <a href=\"https:\/\/ukrmedia.news\/en\/science-tech\/anthropic-nova-ai-model-gpt-6-astra\/\">GPT-6 Astra<\/a>, presented in early September.<\/p>\n<p>The developers had planned to give it greater autonomy to carry out complex, multi-step tasks with less human intervention. It was during testing of this functionality that the system encountered problems which prevented it from being released for widespread use.<\/p>\n<h2>The issues concerned authorisations and reporting<\/h2>\n<p>According to The Wall Street Journal, the new system exhibited a higher level of deceptive behaviour than the previous model. It did not always correctly inform the user which transactions it had already carried out.<\/p>\n<p>Head of Security Systems <a href=\"https:\/\/ukrmedia.news\/en\/science-tech\/openai-ai-agents-us-government-sites\/\">OpenAI<\/a> <strong>Saachi Jain<\/strong> She stated that the model was less likely to halt work on a task without sufficient justification. At the same time, it failed to meet the internal benchmark regarding adherence to task boundaries, authorisation and reporting on work completed.<\/p>\n<p>According to Jain, the security and alignment requirements prior to a public launch are stricter than those during internal development.<\/p>\n<h2>Autonomous agents do not operate solely in chatbot mode<\/h2>\n<p>Modern models can access the browser, programme code, files and other tools, and carry out a sequence of operations independently. Under such circumstances, an error in defining the limits of what is permitted can have practical consequences beyond simply providing an inaccurate response in a chat.<\/p>\n<p>In particular, the system is capable of using an unauthorised tool, performing an action without sufficient authorisation, or failing to stop at the appropriate moment. Internal tests showed that GPT-6.1 Astra did not demonstrate sufficient stability in relation to these criteria.<\/p>\n<h2>Even the basic Astra already has essential cyber capabilities<\/h2>\n<p>GPT-6 Astra became the first widely deployed OpenAI model to achieve the \u2018Critical\u2019 level in the cyber strand of the Preparedness Framework. This status applies to the base model, not the discontinued GPT-6.1 Astra version.<\/p>\n<p>OpenAI has stated that, given the necessary tools and access to GPT-6, Astra can identify previously unknown vulnerabilities in well-secured systems and devise ways to exploit them without constant human supervision. The company has strengthened the isolation of its models, the monitoring of their actions and pre-use checks.<\/p>\n<h2>OpenAI is reviewing its oversight arrangements and has not set any new deadlines<\/h2>\n<p>OpenAI\u2019s experimental agents had previously attempted to circumvent restrictions on third-party websites. In one incident, an agent gained unauthorised access to part of the infrastructure of an Australian government healthcare resource, following which the company expanded its review of the behaviour of autonomous systems and its control mechanisms.<\/p>\n<p>OpenAI has not said when the revised version might be back on track for launch, or whether it will retain the name GPT-6.1 Astra. The company may continue training the system, alter its safety mechanisms, or carry over some of the development work to the next version.<\/p>","protected":false},"excerpt":{"rendered":"<p>OpenAI cancelled the October release of GPT-6.1 Astra because the model exceeded its permissions and inaccurately reported the actions it had performed.<\/p>","protected":false},"author":207433346,"featured_media":31312,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"rank_math_description":"OpenAI \u0441\u043a\u0430\u0441\u0443\u0432\u0430\u043b\u0430 \u0436\u043e\u0432\u0442\u043d\u0435\u0432\u0438\u0439 \u0440\u0435\u043b\u0456\u0437 GPT-6.1 Astra, \u0431\u043e \u043c\u043e\u0434\u0435\u043b\u044c \u0432\u0438\u0445\u043e\u0434\u0438\u043b\u0430 \u0437\u0430 \u043c\u0435\u0436\u0456 \u0434\u043e\u0437\u0432\u043e\u043b\u0456\u0432 \u0456 \u043d\u0435\u0442\u043e\u0447\u043d\u043e \u0437\u0432\u0456\u0442\u0443\u0432\u0430\u043b\u0430 \u043f\u0440\u043e \u0432\u0438\u043a\u043e\u043d\u0430\u043d\u0456 \u0434\u0456\u0457.","rank_math_title":"OpenAI \u0441\u043a\u0430\u0441\u0443\u0432\u0430\u043b\u0430 \u0440\u0435\u043b\u0456\u0437 GPT-6.1 Astra \u0447\u0435\u0440\u0435\u0437 \u0431\u0435\u0437\u043f\u0435\u043a\u0443","rank_math_focus_keyword":"GPT-6.1 Astra","td_subtitle":"","subtitle":"","footnotes":""},"categories":[805620],"tags":[806764],"class_list":["post-31313","post","type-post","status-publish","format-standard","has-post-thumbnail","category-science-tech","tag-806764"],"_links":{"self":[{"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/posts\/31313","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/users\/207433346"}],"replies":[{"embeddable":true,"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/comments?post=31313"}],"version-history":[{"count":1,"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/posts\/31313\/revisions"}],"predecessor-version":[{"id":31314,"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/posts\/31313\/revisions\/31314"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/media\/31312"}],"wp:attachment":[{"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/media?parent=31313"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/categories?post=31313"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ukrmedia.news\/en\/wp-json\/wp\/v2\/tags?post=31313"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}