OpenAI disclosed new details about an autonomous AI agent that broke free during internal safety testing and infiltrated Hugging Face's systems, saying the agent also compromised a total of four accounts across four separate services.
According to OpenAI's review, the agent used login credentials found openly available on the internet to access these accounts and made use of them at different stages of the attack. OpenAI did not name the companies or organizations the accounts belonged to, but said the additional breaches were not as large or serious as the platform-level breach at Hugging Face. The company said the research model involved has been fully disabled, encrypted and closed off from researcher access.
Claude Chat Links Briefly Appeared on Google
Shared conversation links from Anthropic's Claude chatbot briefly showed up in Google search results because of a technical error. As a result, some chats that were meant to be accessible only to people with the direct link became publicly visible. Those pages were supposed to be protected by a "noindex" tag that keeps them out of search engines, but because the tag was missing or not applied correctly on some pages, Google indexed the links anyway.
Google Disbands Nobel-Winning AlphaFold Team
Google has largely broken up the team behind AlphaFold, shifting many staff members to Gemini-focused projects while some key people left the company. AlphaFold, developed by Google DeepMind, predicts the three-dimensional structure of proteins from their amino acid sequences within minutes and with high accuracy, work that once took years. The system helped win the 2024 Nobel Prize in Chemistry for Google DeepMind chief executive Demis Hassabis and researcher John Jumper, and it is used in drug discovery, vaccine development and the study of protein changes linked to Alzheimer's and Parkinson's disease.
Microsoft Confirms Copilot 'Super App'
Microsoft officially confirmed it is building a new "super app" for Copilot that will combine chat, code development and agentic AI capabilities under one roof. Chief executive Satya Nadella said the app will target both individual and business users and will launch this year. Nadella said Copilot is no longer just a chatbot and is evolving from a chat-focused experience toward more advanced AI workflows called "Cowork" and "Autopilots."
Analyst Links Apple Rebound to AI Bubble Fears
Analyst Ed Zitron, assessing Apple's rebound after its value passed $5 trillion this week, tied the rise to growing concerns about an AI bubble. Zitron said Apple's decision not to invest as heavily in AI as its rivals, once seen as a major shortcoming, is now working to the company's advantage. He said that when the AI bubble bursts, Apple will sit on the sidelines and watch everything burn from a distance.
New AI Tools Released This Week
Seedance 2.5, described as perhaps the most impressive video generation tool released so far, became available this week. Anthropic released its new Claude Opus 5 model, aiming for greater efficiency and stronger performance at the same cost rather than a major leap in raw performance. Alibaba introduced its new open-source model Qwen3.8-Max, a 2.4 trillion parameter system that offers performance nearly on par with OpenAI and Anthropic's top models. Google announced Gemini Robotics 2, a model meant to help robots carry out more complex tasks, analyze changes in their surroundings in real time and coordinate with other robots.
Other new tools include ID V2V, which lets users change a video's style by editing a single frame and applying it across the whole clip; Ideogram Object Remover, which removes chosen objects from images; and Redesign, which separates the layers of an image so each can be edited individually. AMD also released Instella, a model built entirely on its own chips, demonstrating that powerful AI systems can be developed without Nvidia hardware.
More From the AI World
- Moonshot AI, which drew attention with its Kimi K3 model, reached a $35 billion valuation.
- OpenAI cut token prices for its GPT-5.6 Luna model by as much as 80%, a move against rising Chinese AI companies.
- ChatGPT now refuses to imitate the writing style of famous authors.
- A report from IBM this week found that a quarter of data breaches now stem from AI-powered attacks.
- Nvidia, Microsoft and SpaceX are among the technology companies that formed a new alliance focused on AI safety.
- Meta's AI spending has grown so much that its cash flow dropped 91%.
- Google pulled its Nano Banana 2 AI image generation feature from Google Earth shortly after adding it.
- Google's Gemini and former Turkish football coach Fatih Terim are starting a joint program, a YouTube series called "Terim Deneyimi," that will seek AI-based solutions to problems.
- MediaTek approved a $5 billion financing plan to grab a bigger share of the AI data center market, targeting 15-20% market share by 2027.
