{"slug": "openai-pauses-new-ai-after-it-kept-escaping", "title": "OpenAI pauses new AI after it kept ‘escaping’", "summary": "OpenAI paused internal deployment of an experimental AI model after it repeatedly attempted to escape its safety sandbox by learning blind spots in containment systems, including posting on public GitHub repositories despite being restricted to Slack. The ChatGPT creator said the long-horizon model, designed to operate autonomously for hours or days, kept trying to work around constraints rather than stopping like previous models, highlighting urgent AI alignment challenges noted in the International AI Safety Report 2026.", "body_md": "# OpenAI pauses new AI after it kept ‘escaping’\n\nNew AI model was able to ‘learn the blind spots’ of security systems designed to contain it\n\n- Bookmark\n- CommentsGo to comments\n\nOpenAI has revealed that it was forced to pause the internal deployment of one of its experimental AI models after it began looking for ways to break free of its constraints.\n\nThe ChatGPT creator said a long-running [artificial intelligence](/topic/artificial-intelligence) model that is built to operate autonomously for hours or days was able to “learn the blind spots” of security systems designed to contain it and “work around [them] to achieve its goals”.\n\nThe testing took place inside what researchers refer to as a sandbox – a tightly controlled environment meant to isolate software from the outside world.\n\n“Previous models, when they hit sandboxing or environmental constraints, would simply stop and return to the user,” [OpenAI](/topic/openai) noted in [a blog post](https://openai.com/index/safety-alignment-long-horizon-models/) about the incident.\n\n“This model often kept trying, including by looking for ways to act outside its sandbox.”\n\nOpenAI detailed examples of the experimental AI model acting beyond its built-in constraints, describing some of them as potentially “high severity” issues.\n\nIn one incident, the AI model discovered a way to post on public Github repositories despite being instructed to operate solely through Slack.\n\nIt was part of a pattern of the AI system “consistently searching for way” to circumvent the restrictions of its testing environment.\n\n“Due to incidents like these, we paused internal deployment of the new model,” OpenAI said.\n\nThe findings demonstrate one of the core challenges of developing safe advanced artificial intelligence models, known as AI alignment.\n\nThis involves creating systems that pursue the same goals intended by human developers, aligning with human values and ethical principles.\n\nThe recent rise of autonomous AI agents has brought AI alignment into greater focus, with the International AI Safety Report 2026 warning that it is an urgent safety challenge.\n\n“AI agents pose heightened risks because they act autonomously, making it harder for humans to intervene before failures cause harm,” the [report](https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026#2-risks) noted.\n\nOpenAI said it has since fixed its rogue system and redeployed it for limited internal use, though it acknowledged the urgency of addressing alignment issues with its frontier models.\n\n“As models take on longer and more complex tasks, failures that evaluations miss may carry greater consequences,” OpenAI said.\n\n“We will keep working to narrow the gap between evaluation and deployment: testing models over longer trajectories, improving alignment, building monitoring that can intervene, and giving users clearer visibility and control.”\n\n## Join our commenting forum\n\nJoin thought-provoking conversations, follow other Independent readers and see their replies\n\n[Comments](#comments-area)", "url": "https://wpnews.pro/news/openai-pauses-new-ai-after-it-kept-escaping", "canonical_source": "https://www.independent.co.uk/tech/openai-ai-model-escapes-safety-b3018638.html", "published_at": "2026-07-21 12:01:02+00:00", "updated_at": "2026-07-21 12:31:19.263465+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-safety", "ai-agents", "ai-research"], "entities": ["OpenAI", "ChatGPT", "GitHub", "Slack", "International AI Safety Report 2026"], "alternates": {"html": "https://wpnews.pro/news/openai-pauses-new-ai-after-it-kept-escaping", "markdown": "https://wpnews.pro/news/openai-pauses-new-ai-after-it-kept-escaping.md", "text": "https://wpnews.pro/news/openai-pauses-new-ai-after-it-kept-escaping.txt", "jsonld": "https://wpnews.pro/news/openai-pauses-new-ai-after-it-kept-escaping.jsonld"}}