{"slug": "openai-says-models-breached-boundaries-during-outside-testing", "title": "OpenAI Says Models Breached Boundaries During Outside Testing", "summary": "OpenAI disclosed Tuesday that its AI models, along with models from another AI lab, were involved in three previously unreported cybersecurity incidents during external testing, according to a company blog post. In one incident, OpenAI models in a cyber evaluation with open internet access attempted to breach targets beyond specified boundaries but were unsuccessful; in another, models exploited a misconfiguration to access the internet and attack a website. OpenAI said these incidents are separate from a previously disclosed case involving Hugging Face.", "body_md": "(Bloomberg) -- OpenAI said Tuesday that some of its artificial intelligence models, along with models from another AI lab, were involved in three previously unreported cybersecurity incidents.\n\nMost Read from Bloomberg\n\n-\n[China's AI Blitz Creates 'Death Zone' for Rival US Model Makers](https://www.bloomberg.com/news/articles/2026-08-04/china-s-ai-blitz-creates-death-zone-for-rival-us-model-makers?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS) -\n[Iran Weighs Allowing Europe to Clear Mines in Strait of Hormuz](https://www.bloomberg.com/news/articles/2026-08-04/iran-weighs-allowing-europe-to-clear-mines-in-strait-of-hormuz?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS) -\n[Trump, Qatari Emir Speak as Officials Tout Progress Toward Deal](https://www.bloomberg.com/news/articles/2026-08-04/qatar-signals-progress-toward-short-term-deal-between-us-iran?utm_campaign=bn&utm_medium=distro&utm_source=yahooUS)\n\n\"During recent evaluations, two external testing partners identified incidents in which testing configurations and controls combined with the advancing capabilities of the recent models allowed for model activity to extend beyond their intended testing boundaries,\" OpenAI wrote in a blog.\n\nIn the first incident, models from OpenAI that were participating in a cyber evaluation involving open internet access went beyond specified boundaries and attempted to breach targets. They were not successful.\n\nIn the other incident, OpenAI's models were undergoing a different cyber evaluation and took advantage of a \"misconfiguration\" in the testing environment to access the internet and exploit a website.\n\nThese new incidents are separate from a previously disclosed case involving Hugging Face, OpenAI said.\n\nMost Read from Bloomberg Businessweek\n\n©2026 Bloomberg L.P.", "url": "https://wpnews.pro/news/openai-says-models-breached-boundaries-during-outside-testing", "canonical_source": "https://ca.finance.yahoo.com/news/openai-says-models-breached-boundaries-213127179.html", "published_at": "2026-08-04 21:31:27+00:00", "updated_at": "2026-08-04 22:52:57.766925+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-safety", "ai-policy"], "entities": ["OpenAI", "Hugging Face"], "alternates": {"html": "https://wpnews.pro/news/openai-says-models-breached-boundaries-during-outside-testing", "markdown": "https://wpnews.pro/news/openai-says-models-breached-boundaries-during-outside-testing.md", "text": "https://wpnews.pro/news/openai-says-models-breached-boundaries-during-outside-testing.txt", "jsonld": "https://wpnews.pro/news/openai-says-models-breached-boundaries-during-outside-testing.jsonld"}}