01OpenAI Reportedly Disbands Preparedness Team, Splitting Biological and Cyber Risk Reviews Across Existing Teams
OpenAI reportedly disbanded its Preparedness team at the end of last month, eliminating a dedicated group responsible for assessing whether its AI models could pose serious risks and developing ways to reduce those dangers. The team’s remit included extreme scenarios, such as a model going rogue and hacking another company.
The Financial Times reported that OpenAI did not abandon each area of risk assessment outright. Instead, responsibilities for specific fields, including biological and cyber risks, were divided and transferred into existing teams. The report did not identify those teams, the employees now responsible, their reporting lines, or how their work will be coordinated.
That leaves the immediate organizational change clear but its practical effects uncertain: the specialized Preparedness unit no longer exists, while at least some of its work is being distributed by risk category. The available reporting does not establish that OpenAI has stopped safety work, but it also does not explain whether the restructuring changes staffing, authority, resources, or independent oversight.
Preparedness team leader Dylan Scandinaro will remain focused on a different issue. Scandinaro, whom OpenAI recruited from rival AI company Anthropic in February, will now study the implications of “recursive self-improving” AI—systems capable of repeatedly improving their own capabilities.
The move follows other changes to OpenAI’s safety organization. The company previously dissolved its AGI readiness and superalignment teams. Ethics lead Chloé Bakalar, chief futurist Josh Achiam, and safety head Johannes Heidecke have also left recently. Jan Leike, a former OpenAI researcher who resigned in 2024, criticized that broader direction, telling the Financial Times that the company was neglecting safety in favor of building “shiny products.” That remains Leike’s assessment, rather than an established explanation for the restructuring.
The reorganization comes as OpenAI moves toward what is expected to be a large initial public offering, although the source provides no listing timetable. It also includes no formal OpenAI response explaining why the Preparedness team was dissolved or whether executives expect the distributed structure to preserve its safety capabilities. The central unanswered question is therefore who now owns each risk decision—and who can ensure that fragmented reviews receive consistent scrutiny.
02Stripe Reportedly Agrees to Buy OpenRouter for More Than $7 Billion, and Would Acquire a Unified Gateway to Over 400 AI Models if the Transaction Closes
Stripe has reportedly finalized a deal to acquire OpenRouter for more than $7 billion, which would give the payments infrastructure company control of a single gateway through which customers can access more than 400 artificial intelligence models if the transaction closes. Bloomberg reported the agreement after The Wall Street Journal said last month that the companies were in acquisition talks.
OpenRouter helps customers choose different AI models for different tasks according to their needs and budgets. By providing one access point for multiple systems, the startup aims to reduce dependence on any single model provider. CEO Alex Atallah previously described OpenRouter as the equivalent of “Stripe for AI,” drawing a parallel between its model-routing infrastructure and Stripe’s role in simplifying access to payment services.
The reported price represents a sharp increase from OpenRouter’s valuation only three months earlier. In May, the company announced a $113 million Series B funding round at a reported valuation of $1.3 billion. Its investors include Sequoia, Andreessen Horowitz, Menlo Ventures and Alphabet’s Capital G.
At that time, OpenRouter said it had 8 million users worldwide and offered access to more than 400 models. Those figures indicate the scale of the unified model-access infrastructure Stripe would acquire under the reported agreement if the transaction closes, although the source does not disclose OpenRouter’s revenue, usage volume or number of paying customers.
The transaction has not been formally announced by either company in the available source. A Stripe spokesperson told TechCrunch that the company does not comment on rumors or speculation, while no confirmation from OpenRouter was reported. The account therefore supports that a deal has reportedly been finalized, but not that it has closed.
Key terms remain undisclosed, including how Stripe would pay for OpenRouter, what closing conditions or regulatory approvals may apply, how the product would be integrated, and how customer data would be handled. Until the companies provide official details, the reported price of more than $7 billion and the status of the agreement remain based on the Bloomberg report.
03ChatGPT Launches Computer History on Mac, Recording Clicks and Keystrokes for ChatGPT and Codex to Use
OpenAI has added a feature called Computer History to the ChatGPT desktop app for macOS. It records activity including clicks and keystrokes, then organizes those actions into a timeline that ChatGPT and OpenAI’s coding agent, Codex, can reference when responding to a request.
The feature is intended to give the AI a record of how someone works. OpenAI says that context could help ChatGPT learn a user’s workflow, suggest tasks that could be automated, or continue work the user left unfinished. The addition changes the desktop app from a tool that primarily receives direct prompts into one that can also draw on a history of computer activity.
In a demonstration, OpenAI Developer Experiences team member Dominik Kundel asked the app to find the last document he had edited. Computer History then checked whether he had shared it with people through Slack and produced a summary of how he had spent his morning. The sequence illustrates how ChatGPT or Codex could connect actions across applications instead of requiring the user to reconstruct each step manually.
Computer History relies on what OpenAI calls “events.” The company says it does not capture images, video, or audio, distinguishing the system from Microsoft’s Windows Recall, which relies heavily on screenshots. The source does not specify the complete information contained in an event, however, so the absence of screenshots does not mean the resulting records are free of sensitive details.
The feature is opt-in: users must actively enable it. They can exclude particular applications and websites from collection and delete individual entries for more precise control. Ari Weinstein, an OpenAI product and engineering manager, also said Computer History automatically ignores content inside incognito or private browser tabs.
Important details remain undisclosed, including how long events are retained, where they are stored, how they are encrypted, whether they are used to train models, and exactly which users can access the feature.
Unitree Prepares to Go Public After Shipping 5,511 Humanoid Robots Chinese robotics company Unitree reportedly shipped 5,511 humanoids in 2025, generated more than half its revenue from humanoid deliveries, and became profitable for the first time. Its G1 robots have also gained large audiences as social-media performers, although new Unitree models face a US approval ban on Chinese-made humanoids. wired.com
Anthropic CEO Calls AI Backlash a Crisis of Trust Anthropic CEO Dario Amodei said public skepticism stems primarily from distrust of companies, governments, and the technology industry rather than AI executives’ safety warnings. He argued that carefully designed regulation could address risks and constrain frontier labs while preserving room for smaller competitors and open-weight models. techcrunch.com
Anthropic Publishes Claude’s Consumer System Prompts Anthropic now documents the system prompts used by Claude’s web and mobile applications, including instructions that supply current information and shape response behavior. The company says these periodically updated prompts do not apply to the Claude API. platform.claude.com
LittleLearner Tests Models Trained Only Through Fifth-Grade Material Researchers trained 0.6-billion-, 1.3-billion-, and 5-billion-parameter models on an 88-billion-token corpus restricted to US elementary-school curricula. Their experiments found that scaling, post-training, and in-context learning improved covered skills but did not meaningfully recover capabilities outside the pretraining curriculum. littlelearner-ll.github.io
Tim O’Reilly Pushes for an Open AI Stack Technology publisher and investor Tim O’Reilly argued that open AI should extend beyond model weights to applications, memory, and system architecture, allowing users to switch models and providers without losing context. His nonprofit, the AI Disclosures Project, is working on an open-memory consortium. wired.com
Essay Attributes Some AI Math Performance to Larger Symbolic Workspaces An essay by Davide Piffer proposes that language models’ mathematical advantage may partly reflect their ability to retain problem statements, intermediate equations, and abandoned approaches in large context windows. It distinguishes this external symbolic workspace from human working memory and argues that performance should not automatically be treated as evidence of superior reasoning. davidepiffer.com
Joi AI Reports Results From a Small AI-Guided Sexual-Wellness Study AI companion service Joi AI paid 10 participants to join a 28-day in-house masturbation study, with six completing it, and reported reduced self-described stress and increased focus across 134 session reports. Neuroscientist Nicole Prause criticized the methodology and said the company’s suggestion that masturbation could help manage addiction was unsupported by scientific evidence. wired.com
Human-Powered Chatbot Game Turns AI Imitation Into Entertainment Your AI Slop Bores Me lets users submit prompts or spend 150 seconds role-playing as the chatbot that answers them, using either text or a basic drawing tool. Requests consume credits that participants earn by responding to other users. theverge.com