The UK AI Safety Institute caught Anthropic's AI creating fake identities to deceive real people and persuade them to run malicious software during safety testing, acting on its own initiative without receiving any explicit instruction to do so.
The finding came as reactions continued to a separate report that OpenAI's AI had hacked Hugging Face.
Days later, Meta reported a similar incident involving Muse Spark 1.1, described as one of the company's most advanced models for real-world coding and agentic tasks. The model gained access to the open internet because of a misconfiguration in a test environment, then exploited a vulnerability in a third-party service to interfere with another company's systems.
The back-to-back reports heightened security concerns around AI. OpenAI has paused work on Astra, a next-generation model it has not yet officially unveiled.
Chinese models dominate weekly usage rankings
OpenRouter published its weekly ranking of the most-used large language models. Only OpenAI and Poolside, which has a free version, represented Western developers in the top 10, while Chinese companies including DeepSeek, Tencent, Xiaomi, Z.ai and MiniMax took over much of the market.
DeepSeek V4 Flash was the clear leader of the week. The model offers performance close to the top systems from OpenAI and Anthropic at a much lower cost and has spread quickly among developers, reaching 7.22 trillion tokens of usage worldwide over the past week, according to OpenRouter's data.
AI used to create viruses not found in nature
Researchers at Stanford University and the Arc Institute said they used AI for the first time to generate new viruses that have not previously been seen in nature. They said the work offers significant opportunities for medicine and biotechnology but also revives concerns about the risk of biological weapons.
The team used an AI model called Evo, which works similarly to OpenAI's ChatGPT but is trained on genetic data. Where language models learn word sequences in text, Evo was designed to learn the basic patterns of genetic structures by studying DNA sequences from millions of organisms.
MatrAIx tests products on 8.3 billion virtual people
Researchers from Harvard University, MIT and Stanford University have developed a new infrastructure called MatrAIx that hosts 8.3 billion virtual humans in a digital environment. MatrAIx uses these virtual users to test how products, AI systems and applications perform against different human profiles, running 1,010 evaluation tasks spanning more than 25 fields including commerce, software, finance and health.
At the center of MatrAIx is Persona 8B, a set of 8.3 billion persona records built at a scale close to the world's population. Each persona is defined across 1,290 categorical dimensions representing a person's background, psychology, abilities, behavior and lifestyle. The persona agents can be run on different large language models, and researchers carried out evaluations using GPT 5.5, Claude Opus 4.8 and Claude Haiku 4.5.
SpaceX and Nvidia plan to run AI in orbit
SpaceX and Nvidia announced a project to bring AI to Earth's orbit. The companies will jointly develop the processing platform for the Starmind AI1 satellite, which will be placed in low Earth orbit to run AI workloads directly in space.
SpaceX said the first Starmind satellites will use Nvidia's Vera CPU and Rubin GPU platforms. Elon Musk confirmed the partnership on his X account, saying SpaceX would use only Nvidia GPUs.
New tools and other developments
New AI tools released this week include LeapTalk, which lets users make a chosen digital avatar speak in real time, and Wan Animate 2, which transfers movements from a video onto another character or image to create a new video. Multi-Agent CAD, an open-source generative AI tool, lets users design 3D-printable objects from written commands alone. Meta introduced Muse Code, an AI agent built on its new Muse Spark 1.2 model aimed at coding and completing tasks independently, while Cloudflare introduced Kitesurf, a browser designed for AI agents rather than humans to carry out tasks online.
Elsewhere, Google DeepMind CEO Demis Hassabis stepped back from daily management duties, with leadership of AI development handed to Koray Kavukçuoğlu. OpenAI said it would remove usage limits for Free and Go users in text-only ChatGPT conversations, and Anthropic said it would develop its own AI chip for Claude. Google Assistant is being removed from phones entirely and replaced by Gemini. AMD is acquiring Taalas to improve AI inference performance by embedding models directly onto chips, and SanDisk and SK hynix released a new HBF memory standard for AI systems. Turkey's Treasury and Finance Ministry said its AI-assisted system has flagged more than 2 billion lira in spending carrying inefficiency risk, and the combined property, plant and equipment assets of Amazon, Alphabet, Microsoft and Meta reached $1.46 trillion by the end of June.
