OpenAI has released a beta App Directory inside ChatGPT that lets users browse and use third party apps directly within a conversation. Apps can be launched from the sidebar or invoked mid chat by mentioning them by name, so external tools can be brought into the flow of a single session to handle actions such as generating slides, searching listings, or completing multi step workflows that begin as natural language prompts. Developers can now submit apps for review through the OpenAI developer platform, and approved apps will roll out gradually in the new year as the company positions ChatGPT as a hub for specialized capabilities.
Google is testing an experimental Artificial Intelligence agent called CC that delivers a personalized “Your Day Ahead” briefing each morning by drawing on Gmail, Google Calendar, Google Drive, and the web to summarize schedules, flag important updates, and surface tasks like upcoming appointments or pending bills. CC can also prepare email drafts and calendar links so users can quickly act on those briefings, and it is built on the Gemini model so that it can learn from user feedback over time. CC is launching in early access to paid subscribers over 18 in the US and Canada, and Google has opened a waitlist as it explores more proactive, assistant style services.
In funding news, a16z announced a $16M Series A investment in Oboe, which turns a single prompt into a structured, chaptered course containing text, audio, visuals, quizzes, and flashcards so that users get a guided learning path instead of an unstructured collection of resources. In mobility, code in Waymo’s mobile app indicates the company is experimenting with a Gemini powered ride assistant for its robotaxis that can answer rider questions, personalize greetings, and adjust in cabin settings like temperature and lighting while explicitly avoiding commentary on real time driving behavior or vehicle decisions. At the same time, OpenAI has warned that agentic Artificial Intelligence browsers may always face persistent prompt injection risks, and says it is pursuing layered security controls, faster incident response cycles, and reinforcement learning based automated attackers that simulate real world threats to find vulnerabilities first.
For alignment researchers, Anthropic has open sourced Bloom, a framework for generating behavioral evaluations of frontier Artificial Intelligence models. Bloom lets researchers define a specific behavior and automatically generate diverse scenarios to measure how often that behavior appears, addressing the way fixed test sets can quickly become outdated as models evolve. The system runs a multi stage pipeline that creates scenarios, executes multi turn rollouts, and scores outputs using judge models to produce metrics such as elicitation rate, with configuration seeds used to keep evaluations reproducible while still allowing rapid iteration.
