Post Thumbnail

Closed Microsoft and Google repositories found accessible to Copilot

Israeli company Lasso, specializing in cybersecurity in the field of generative artificial intelligence, discovered a serious security issue: data that was accessible on the internet even for a short time continues to exist in the memory of AI chatbots even after being deleted or switched to private mode.

The scale of the problem turned out to be impressive: more than 20,000 GitHub repositories, which are now private, are still accessible through Microsoft Copilot. This vulnerability affects over 16,000 organizations, including tech giants such as Google, IBM, PayPal, Tencent, and Microsoft itself.

The story began when Lasso discovered its own repository, which was accidentally made public for a short time, in Copilot’s responses. “If you browse the web, you won’t see this data. But any person in the world can ask Copilot the right question and get this information,” explains Lasso co-founder Ofir Dror.

The research showed that the problem is related to the caching mechanism of Microsoft’s Bing search engine. Lasso analyzed a list of repositories that were public at some point in 2024 and were then deleted or switched to private mode. It turned out that data from these repositories is still accessible through Copilot.

This discovery raises serious questions about data security in the era of generative AI. Even a brief disclosure of confidential information can lead to long-term consequences, as AI systems index and store this data, making it potentially accessible to any user who knows the right questions.

The situation is particularly notable because it affects the world’s largest technology companies, including the developer of the tool itself – Microsoft. This demonstrates that even organizations with the highest level of technical expertise can face new security challenges created by artificial intelligence systems.

Autor: AIvengo
For 5 years I have been working with machine learning and artificial intelligence. And this field never ceases to amaze, inspire and interest me.

Latest News

First Google Veo 3 advertisement shown to millions during NBA finals

Millions of NBA finals viewers witnessed a completely new stage in creative evolution. Fully computer algorithm-generated advertisement for betting platform Kalshi, created using Google Veo 3.

Chinese platform QiMeng creates processors at Intel 486 and Arm level

Chinese scientists developed a new AI platform capable of independently designing processors at the level of human experts. Researchers from the State Laboratory for Processor Development and the Intelligent Software Research Center presented an open-source project called QiMeng.

Meta AI turns private AI chats into public posts without knowledge

Meta AI app turned out to be a real catastrophe for user privacy. Turning their private conversations with artificial intelligence into public content. Imagine a modern horror movie: your entire query history became publicly accessible, and you didn't even suspect it.

Google released Gemini 2.5 Flash-Lite with 1 million token context

Google introduced the Gemini 2.5 Flash-Lite model. And it becomes a real breakthrough in price-performance ratio, opening new horizons of advanced technology accessibility.

Anthropic created 17 virtual worlds for testing AI sabotage

Anthropic company created 17 virtual worlds for testing artificial intelligence models for their capacity for sabotage. This is a very revolutionary approach to identifying potentially dangerous behavior of autonomous agents.