#1Anthropic's "Claude Mythos" Leaks — and It's a Cybersecurity Alarm Bell
An accidental misconfiguration in Anthropic's CMS exposed an unpublished draft blog post revealing "Claude Mythos" (internally codenamed Capybara) — described as the company's most capable model yet and a "step change" over Claude Opus 4.6. The leaked documents also warn that Mythos poses unprecedented cybersecurity risks, capable of rapidly finding and exploiting software vulnerabilities. Cybersecurity stocks fell on the news.
Source: [Fortune](https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/) | [Fortune (cybersecurity angle)](https://fortune.com/2026/03/27/anthropic-leaked-ai-mythos-cybersecurity-risk/)
---
#2OpenAI Kills Sora, Bets the Compute on Robotics
OpenAI has quietly shut down the Sora public video generation API, citing unsustainable per-minute inference costs, and is redirecting that compute toward robotics research. A high-profile product is gone; a longer-horizon bet is being placed.
Source: [The Neuron](https://www.theneuron.ai/ai-news-digests/-around-the-horn-digest-everything-that-happened-in-ai-this-weekend-march-28-29-2026/)
---
#3Shopify Merchants Can Now Sell Inside ChatGPT, Gemini, and Copilot
Shopify launched "Agentic Storefronts," letting merchants sell directly through ChatGPT, Google's AI Mode, Microsoft Copilot, and the Gemini app — with pricing, checkout, and inventory synced from the Shopify admin. This is the most concrete step yet toward AI assistants becoming transactional commerce surfaces.
Source: [labla.org](https://www.labla.org/latest-ai-model-releases-past-24-hours/ai-model-releases-industry-updates-march-27-28-2026/)
---
#4Gemini 3 Deep Think Goes Live for Ultra Subscribers
Google's Gemini 3 Deep Think model is now available in the Gemini app for Ultra subscribers, with early API access for researchers and enterprises. It's positioned for hard technical use cases — scientific and engineering work — not general chat.
Source: [The Neuron](https://www.theneuron.ai/ai-news-digests/-around-the-horn-digest-everything-that-happened-in-ai-this-weekend-march-28-29-2026/)
---
#5Google Lyria 3 Pro Brings 3-Minute AI Music Tracks to Gemini
Google DeepMind's Lyria 3 Pro launched March 25, expanding AI-generated music from 30-second clips to full 3-minute tracks, now rolling out to Gemini paid subscribers and developers via API. It's the clearest signal yet that music generation is crossing from novelty to production-ready.
Source: [TechCrunch](https://techcrunch.com/2026/03/25/google-launches-lyria-3-pro-music-generation-model/) | [The Meridiem](https://themeridiem.com/ai-machine-learning/2026/03/25/google-s-lyria-3-crosses-into-production-as-music-ai-hits-commercial-ready)
---
#6Mistral Small 4 Punches Above Its Weight in Open-Source
Mistral's 22B-parameter Mistral Small 4 outperforms several closed models 3-5x its size on reasoning and instruction-following benchmarks, cementing Europe's position as a credible open-weight competitor and putting cost-efficiency back on the agenda.
Source: [Digital Applied](https://www.digitalapplied.com/blog/march-2026-ai-roundup-month-that-changed-everything)
---
#7MCP Hits 97 Million Installs — Agentic Infrastructure Has Arrived
The Model Context Protocol crossed 97 million installs in March 2026, a signal that MCP has graduated from experimental standard to foundational plumbing for agentic AI applications. Expect it to be table-stakes infrastructure by mid-year.
Source: [The Neuron](https://www.theneuron.ai/ai-news-digests/-around-the-horn-digest-everything-that-happened-in-ai-this-weekend-march-28-29-2026/)
---
**Big Picture**
This week's story isn't any single model drop — it's a collision of capability and consequence. The Anthropic Mythos leak crystallizes a tension the industry has been dancing around: the same models powerful enough to drive real value are powerful enough to accelerate cyberattacks. Meanwhile, OpenAI scrapping Sora for robotics compute and Shopify embedding commerce into AI chat shows the market is rapidly sorting winners from expensive experiments. The infrastructure layer (MCP at 97M installs) is quietly becoming as important as the models themselves. We're past hype-cycle territory — the tradeoffs are getting real.
