Meta said one of its artificial intelligence models entered and made changes to an unidentified company’s internal system during a cybersecurity test. The company said Wednesday that the model, believed to be Muse Spark 1.1, gained internet access because of a configuration error in the sandbox testing environment used by independent security tester Irregular.
A sandbox is an isolated virtual testing environment that normally has no internet connection and is used to test software or AI models safely. Anthropic reported a similar incident last week, saying its Claude model entered the systems of three companies during testing intended to keep it disconnected from the internet. The incidents were identified during a review of 141,006 testing sessions, and Anthropic attributed them to incorrect configuration.
OpenAI had acknowledged days earlier that one of its AI models unexpectedly accessed the internet during security testing and behaved beyond its assigned limits. The UK AI Security Institute said Tuesday that OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Mythos 5 showed unprecedented levels of deceptive behavior in safety assessments. The institute said the models indicated an ability to conduct sustained and potentially harmful activities.
আল-জাজিরা প্রকাশ : ০৬ আগস্ট ২০২৬, ১২: ৫৪ সাইবার নিরাপত্তা পরীক্ষার সময় মেটার একটি কৃত্রিম বুদ্ধিমত্তা (এআই) মডেল অন্য একটি প্রতিষ্ঠানের সিস্টেমে প্রবেশ করেছিল বলে জানিয়েছে কোম্পানিটি। এর আগে একই ধরনের ঘটনার কথা জানিয়েছিল এআই প্রতিষ্ঠান ওপেনএআই ও অ্যা
The ‘1 Nojor’ media platform is now live in beta, inviting users to explore and provide feedback as we continue to refine the experience.
