Microsoft is adding new Copilot features to Word that target professionals in legal, finance, and compliance.
Copilot can now track changes at the word level, so edits stay transparent and easy to review. Users can also manage comments directly in the text, insert tables of contents, and set up headers and footers with dynamic fields like page numbers.
For multi-step edits, Copilot now shows what it's working on in real time. Microsoft says the features run on "Work IQ," a layer that adapts responses based on the user and their organization. Data stays within Microsoft 365's existing security boundaries. For now, the new features are only available on Windows desktop through the Office Insiders Beta Channel's Frontier program. Web and Mac support will follow.
OpenAI's GPT-5.4 Pro model has apparently solved Erdős open math problem #1196. The model reportedly found the solution in about 80 minutes and prepared it as a LaTeX paper in another 30. Formal verification is underway.
Mathematician Terence Tao commented in the Erdős Problems forum that the work reveals a previously undescribed connection between the anatomy of integers and Markov process theory. "That would be a meaningful contribution to the anatomy of integers that goes well beyond the solution of this particular Erdos problem," Tao writes. Kevin Barreto, who says he'll soon join OpenAI's AI for Science team, noted in the same forum that the Markov chain technique the model used was a creative step human mathematicians had overlooked despite years of work on the problem.
The discussion is interesting because there's an ongoing debate about whether LLMs can discover new knowledge in mathematics and other disciplines that goes beyond the data points learned during training. This example shows that new, previously undescribed knowledge can also be hidden within already known data points.
OpenAI releases GPT-5.4-Cyber, a model built specifically for defensive cybersecurity
OpenAI releases GPT-5.4-Cyber, a model that has been specifically trained for defensive cyber security. Access remains restricted to verified security experts for the time being.
Google is rolling out "Skills," a new Chrome feature that lets users save frequently used AI prompts and reuse them with a single click. Previously, users had to manually re-enter the same prompt each time, for example, converting recipes to vegan alternatives, to cite one of Google's examples.
With Skills, prompts like these can be saved directly from the chat history and pulled up in Chrome by typing a slash ( / ) or plus sign ( + ) in Gemini. The feature works across multiple tabs. Google also offers a library of ready-made skills for things like product comparisons, meal planning, and gift selection. Users can customize these or build their own from scratch.
According to Google, Skills uses Chrome's existing security and privacy features and asks for permission before performing certain actions like sending emails. The feature is rolling out now on Mac, Windows, and ChromeOS for users with their Chrome language set to English-US.
Claude Mythos can autonomously compromise weakly defended enterprise networks end-to-end
The UK’s AI Safety Institute tested Anthropic’s Claude Mythos Preview for cyber capabilities. For the first time, an AI model autonomously completed a full attack simulation against a corporate network, but the results come with significant caveats.
Anthropic has introduced "routines" for Claude Code - automated processes that can independently fix bugs, review pull requests, or respond to events without needing a user's local machine. Routines are configured once and then run on a schedule, via API call, or in response to GitHub events on Anthropic's web infrastructure. Typical use cases include nightly bug triage, automatic code reviews based on team-specific checklists, porting changes between languages, and checking deployments for errors.
Routines tap into existing repository connections and connectors. The feature is available as a research preview for Pro, Max, Team, and Enterprise plans, with daily limits of 5 to 25 runs depending on the plan. Support for webhook sources beyond GitHub is planned.
Users assign a name, describe the task, select a trigger (schedule, GitHub event, or API), and connect external services like Slack or Asana.
Ukraine captures a Russian position using only drones and ground robots
President Zelenskyy announces a historic first: a Russian position taken entirely by unmanned systems. A CSIS report details how AI is already changing Ukraine’s battlefield and where the limits remain.
AI startup Hiro, which built a personal AI financial advisor, is shutting down. The team is joining OpenAI, though no acquisition price has been disclosed. Hiro's vision was a "personal AI CFO" (Chief Financial Officer).
Hiro let users input their salary, debts, and monthly expenses, then calculated financial scenarios and explained the results. The company says it helped customers manage more than one billion dollars in assets.
The deal has all the hallmarks of an acqui-hire, an acquisition aimed at the team, not the product. Hiro has stopped accepting new sign-ups and will shut down on April 20, 2026. Users can export their data until May 13, 2026, after which all personal information will be deleted. No user data will transfer to OpenAI.
OpenAI's leaked memo says new "Spud" model will make all its products "significantly better"
An internal OpenAI memo lays out five strategic priorities for the enterprise business – including a new model codenamed “Spud,” a platform play for AI agents, and the blunt accusation that Anthropic is overstating its revenue by 8 billion dollars.