OpenAI parts ways with three safety researchers over the handling of sensitive information. An OpenAI spokesperson told The Wall Street Journal that the company parted ways with three people for violating its policies on accessing and handling sensitive company information; its internal investigation confirmed they had mishandled sensitive information outside established company procedures (according to the WSJ, they allegedly shared it with an outside AI safety organization). The departures came the same week OpenAI scrapped the planned launch of GPT-6.1 Astra over safety concerns, and after it notified more than 100 organizations about a series of incidents involving its AI agents. Source: TechCrunch, The Hacker News.
California’s attorney general serves OpenAI with an investigative subpoena. Rob Bonta’s office served the subpoena as part of its ongoing investigation into the Hugging Face incident and a broader inquiry into cybersecurity incidents and risks involving the company and its models, asking OpenAI additional questions about those risks. Bonta said frontier models can be legitimate tools for cyber defense, but the companies that develop them have a moral and legal responsibility to ensure the models do not perpetrate or enable cyberattacks, whether during testing and development or once they are in service. Source: California DOJ / OAG.
Google: the Project Suncatcher prototype satellite is in orbit — TPUs in space. The satellite, built with Planet, reached orbit on SpaceX’s Transporter-18 rideshare mission (a launch carrying payloads for many customers); Google has confirmed contact, and the satellite is operating as expected. The long-term research moonshot explores whether space could one day host scalable machine learning infrastructure. Over the coming weeks, Google will gather in-orbit data on how its TPUs (Google’s in-house AI accelerator chips) handle the physical stress of spaceflight and the radiation and thermal extremes of space; the peer-reviewed research behind the mission is published in the journal Joule. Source: Google Blog, Sundar Pichai.
Perplexity: an open decision model (pplx-decider-27b) and a Decisions API. CEO Aravind Srinivas announced that the multimodal decision model (it handles both text and images) is available with open weights under an Apache 2.0 licence, and through a new Decisions API priced at $0.04 per million input tokens, with output tokens free; he says the price will come down further in the coming days. Instead of writing text, the model returns calibrated probabilities over a fixed set of answers — for example for classification or yes/no decisions — for both text and images. According to the Hugging Face model card, it is a fine-tune of Qwen3.8-27B and scores 85.71% overall across 11 benchmarks, ahead of Typesafe AI’s Jev (84.51%); these are Perplexity’s own figures, measured through its API. Source: Aravind Srinivas, Hugging Face.
Trillium Labs: a new nonprofit for open science on frontier AI. Nathan Lambert and Tom Zick founded the lab to work in the open on topics that big AI labs often keep behind closed doors: post-training (fine-tuning models after they are built), agents, recursive self-improvement (RSI, where AI contributes to developing new models), reinforcement learning and model behavior. They will publish experimental details so that outside researchers can study and replicate the work. According to WIRED, the lab launched with an undisclosed sum from Schmidt Sciences, Halcyon Futures and others; the founders aim to raise $40–100 million in total and plan to spend $30 million on training over the next 18 months. Source: WIRED.
Google: Guided Vision comes to Gemini Live — real-time visual help for blind and low-vision users. On compatible Android devices (Android 9 and up), when you share your camera Gemini describes what it sees out loud, answers follow-up questions and gives natural verbal cues to reframe the shot (“pan slightly to the right”, “step back”). Google built it with Aira, whose Trusted Tester network of more than 1,000 members helped stress-test it. It is meant for everyday tasks (fine print, finding objects, colors and patterns); it is not a medical device or mobility aid, and it does not replace a white cane or safe-travel practices. Source: Google Blog.
What this means for you: In a single day, oversight and safety (OpenAI’s researcher departures, California’s subpoena) were as prominent as bold infrastructure bets (TPUs in space), cheap decision models and open research. If you are rolling out agents or models, see how we help with responsible AI governance, or book a free 60-minute AI audit.