Welcome back! Anthropic kept its most powerful security AI in a closed circle for months, and we now know why. OpenAI brought Codex inside the ChatGPT mobile app across every plan. xAI shipped its own coding agent Grok Build at $300 a month with an exodus of researchers behind it. And the Musk-Altman trial finally heads to the jury on Monday.
In today's Generative AI Newsletter:
Mythos x Apple: What did Anthropic's most powerful AI just do to macOS?
Codex mobile: What happens when your coding agent fits in your pocket?
xAI's $300 bet: Can Grok Build catch Claude Code from this far behind?
Musk v Altman: When does the jury decide who wins?
Mythos, available only to a closed circle of researchers under the Claude Mythos Preview programme, just found the first public kernel memory corruption exploit on Apple's M5 chips.
The details:
The find: Two vulnerabilities and several techniques chained together to give an unprivileged local user complete access to macOS Tahoe 26.5.
The AI's role: Mythos identified the bugs and assisted with exploit development, finding new instances faster than humans can.
The trust circle: Anthropic chose not to release Mythos publicly. Apple is among a small number of partners with early access.
The fix: Full technical details on this exploit are yet to be released and will only be shared once Apple ships patches.
This is the most consequential demonstration yet of what frontier AI does when pointed at one of the hardest defensive surfaces in computing.
The kernel of Apple's newest silicon is no joke. But this is good news. Bug discovery has just become much more accessible, which is excellent for defenders who can fix products before they launch to prevent potential exploits.
OpenAI rolled out Codex in preview inside the ChatGPT mobile app for iOS and Android, available to every plan.
The agent keeps running on a laptop or remote host, but live threads, code changes, plugin permissions and model choices now sync to mobile, where users can approve commands or kick off new tasks from anywhere.
The mobile build connects through what OpenAI calls a secure relay layer that does not expose the host machine to the open internet.
It ships alongside two other recent Codex expansions, with desktop background execution added last month and a Chrome extension earlier this month that lets the agent operate inside live browser sessions.
OpenAI's announcement positioned the launch against Anthropic's Remote Control from February and Dispatch from March, claiming Codex now goes further than both.
The release continues OpenAI's push to catch Claude Code, which has held wider business and developer adoption since launch. Let’s see who wins.
SpacexAI shipped Grok Build, an early-beta coding agent and CLI locked to its SuperGrok Heavy tier at $300 a month, with the explicit internal direction to match Claude Code's performance before chasing anything new.
The product is unproven yet and the price is the highest on the market for an agent at this maturity level.
Whether SpacexAI has produced a model that can actually compete with Claude in agentic coding is the only question that matters, and Grok Build's beta status suggests xAI itself is not yet sure.
A $300 floor is a confidence bet on people not currently using the product. If you’re trying it, share your thoughts on it and maybe you’ll get your analysis featured in one of our editions!
After more than two weeks of testimony in Oakland, lawyers for Elon Musk and OpenAI delivered their closing arguments on Thursday. We wanted to share a recap before the case wraps up.
The recap:
Musk filed the suit in February 2024, accusing Altman of betraying OpenAI's nonprofit founding when the company restructured toward a profit-making model.
Of the original 26 claims, 24 have been dismissed or withdrawn over the past two years and two remain.
Musk's lawyer Steven Molo built his closing around testimony from five witnesses calling Altman a liar under oath.
OpenAI's lawyer Sarah Eddy responded that it was Musk whose testimony was contradicted by witnesses.
The jury verdict is advisory and judge Gonzalez Rogers, the same person who presided over the Epic Games vs. Apple case, makes the final liability call.
Musk is seeking billions in disgorgement to fund OpenAI's charitable arm and wants Altman ousted from leadership.
Expect a jury verdict within two weeks, followed by months of judicial deliberation on the remedies phase. The headlines will be loud, but the decisive ruling will not arrive until well after Monday.
Drafted: AI floor plan generator that turns room constraints into downloadable house plans with 3D exterior renders in seconds. Built for homeowners, renovators and builders who would otherwise pay an architect five figures.
Lindy: AI work assistant that runs your inbox, calendar and meetings across Gmail, Outlook, Slack, Notion, HubSpot and Salesforce. It preps you before calls, joins meetings, takes notes and updates your CRM without you opening a second app.
Gumloop: No-code platform for building AI agents that run inside the tools your team already uses, with prebuilt templates for sales, support, data analysis and meeting prep. You call an agent from a Slack channel by tagging @Gumloop the way you would a colleague.
Juicebox: Recruiting tool that turns a single plain-English prompt into a candidate pipeline across 800M+ profiles and 60+ data sources. Evaluates up to 5,000 profiles per search and runs autonomous multi-step outreach around the clock.
Arcade: Production MCP runtime that lets AI agents take real actions across your team's services. Handles OAuth server-side and exposes 7,000+ pre-built integrations so agents act as you without the model ever touching credentials.
Anthropic and Gates pledge $200M for AI public goods: Four-year commitment for health, education and agriculture tools in low and middle-income countries.
Bumble kills the swipe: An AI assistant called Bee replaces left-right matching with richer profile data and ends the women-message-first rule.
Android Auto goes Gemini: Adaptive screens, Immersive 3D navigation, Dolby Atmos and a Gemini agent that handles inbox, takeout and car-spec questions.
AI extracts hidden risks from chest scans: Two RSNA 2025 tools turn routine X-rays and CT scans into opportunistic screening for osteoporosis and cardiovascular risk.
Multiverse raises $70M at $2.1B: Euan Blair's London company rebrands as Europe's AI adoption platform after 50% year-on-year revenue growth.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
🐝 Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🚀 Explore investment opportunities in the future of AI and join our community-backed growth journey.
