Edition: October 3, 2026

Microsoft and Hugging Face introduced ThinkingBox, a benchmark that evaluates AI agents based on the actual changes they make to databases rather than just their generated text. Testing 12 LLM‑driven agents across 507 business workflows, the study found that most agents produce seemingly correct tool calls but often leave incorrect or missing records, highlighting a major reliability gap.

OpenAI safety lead David Robinson resigned, warning that the company’s internal culture is broken and likening its rapid, trial‑and‑error approach to unsafe practices in nuclear or aviation. In an Atlantic essay he argues that without a fundamental cultural shift toward rigorous, redundant safety protocols, increasingly capable AI systems pose growing risks. OpenAI says it is strengthening safeguards, but the departure fuels broader concerns about AI safety governance.

OpenAI safety lead David Robinson has resigned and published an editorial warning that the company’s culture and industry practices are fundamentally broken, beyond mere regulatory fixes. His departure highlights growing internal concerns about the risks of advanced AI models as the sector grapples with safety oversight.

Meta’s new AI assistant Muse can automatically generate detailed profile pages for every person in a user’s life, compiling data from messages, bank accounts and social interactions to suggest ways to improve relationships. Researchers extracted Muse’s internal instructions, revealing how the system structures and stores personal relationship data, raising privacy and ethical concerns about the depth of personal insight AI can gain.

A new wave of AI agents can be accessed directly via text messages, eliminating the need for separate apps. Services like Caddy, Fambot, Folk and the high‑valued Instinct let users schedule appointments, manage family logistics, conduct research and perform real‑world tasks straight from iMessage, SMS or other chat platforms. These agents integrate with existing calendars, email and other services, offering both free beta access and premium subscriptions.

AWS chief Matt Garman announced Amazon has stopped using nondisclosure agreements with government agencies amid growing backlash over data‑center transparency, environmental impact, and community benefits. He defended the industry by disputing myths about water use, electricity costs, pollution, and highlighted Amazon’s $1 billion community contributions, while noting numerous state moratoriums on new data‑center permits.

Meta unveiled Muse Gadgets, an open‑source platform that lets developers attach its personal AI agent to custom hardware like Raspberry Pi or ESP32 boards, providing firmware, a Linux SDK, and community support. The move extends Muse beyond a chatbot, aiming at hobbyists, small businesses, and enterprise customers with new tools such as Muse Home Link and Muse for Small Business.

Capcom announced plans to integrate AI into its game development pipeline, aiming to streamline the creation of large‑scale titles like Resident Evil. While the company maintains it won’t use AI‑generated assets in final games, it sees the technology as a way to speed up production and evolve its RE Engine for next‑gen releases.

Splice CEO Kakul Srivastava warns that AI‑generated emails are eroding genuine conversation, drawing on her experience at tech firms like Flickr and Adobe and her work expanding Splice’s music‑sample platform. She argues that reliance on AI for routine messaging hampers creativity and authentic interaction.

Nvidia announced a $100 price increase for its 7‑year‑old Shield TV Pro, now costing $299 due to rising component costs driven by the AI boom. The hike reflects broader supply‑chain pressures affecting even legacy devices, making the niche streamer less attractive compared to cheaper alternatives.