AI Agents Ship, Mythos Leaks, Apple AI App Store

Microsoft just shipped AI agents that actually do your work. Anthropic accidentally leaked its next model. Apple is building an AI App Store. And Google figured out how to slash AI memory usage by 6x.

Here's what happened in AI this week — and what it means for your business.

Microsoft Copilot Cowork: AI Agents That Execute, Not Just Chat

Microsoft launched Copilot Cowork through its Frontier Program on March 30th, and this one matters. Built on Anthropic's Claude technology integrated directly into Microsoft 365, Cowork doesn't just answer questions — it runs multi-step workflows across your entire Microsoft stack.

Describe the outcome you want, and Cowork creates a plan, reasons across your tools and files, and carries the work forward with visible progress. Think monthly budget reviews, executive briefing prep, calendar management — delegated to an AI agent that actually follows through.

The kicker: Microsoft's new Researcher Critique feature uses multiple AI models in tandem. One model drafts the research, another reviews it for accuracy. The result? A 13.8% improvement on the DRACO benchmark for deep research quality. Capital Group, an early access customer, is already using it for executive review prep.

Why this matters: The "AI copilot" is evolving from a search bar into an execution layer. If your team runs on Microsoft 365, this is the clearest signal yet that AI agents are moving from demos to daily workflows. The multi-model approach — using competing AI systems to check each other's work — is becoming the new standard for enterprise reliability.

Source: Microsoft 365 Blog

Anthropic's Security Lapse Leaks "Mythos" — Its Most Powerful Model Yet

In a twist of irony, the company that positions itself as the safety-first AI lab accidentally left nearly 3,000 unpublished assets in a publicly accessible content management system. Fortune broke the story on March 26th after cybersecurity researcher Alexandre Pauwels at Cambridge flagged the exposure.

Among the leaked materials: details of "Mythos," an unreleased AI model that Anthropic describes internally as representing a "step change" in capabilities. Also exposed were plans for an invite-only CEO event and various internal documents.

Anthropic's response was predictable — "human error in the CMS configuration," not related to their AI tools. The company secured the data after Fortune's notification. But the optics are brutal: the company that lectures the industry on AI safety couldn't secure a basic CMS.

Why this matters: Two takeaways. First, Anthropic's next model sounds significant — if they're calling it a "step change," expect a major capability jump when Mythos ships. Second, this is a reminder that AI safety starts with basic operational security. The biggest AI risks aren't rogue models — they're misconfigured databases and human error. Every company deploying AI should audit their own infrastructure before worrying about sentient machines.

Source: Fortune

Apple's Siri Extensions: The AI App Store Is Coming

Bloomberg's Mark Gurman reported on March 29th that Apple is going far beyond simply plugging ChatGPT into Siri. The company is building Siri Extensions — a framework that lets third-party AI chatbots run inside Siri, with a dedicated App Store section.

Combined with the existing Apple-Google deal giving Apple "complete access" to Gemini for training smaller on-device models via distillation, Apple's strategy is becoming clear: they don't need to build the best AI model. They need to own the distribution layer.

Why this matters: Apple has 2 billion active devices. If they successfully position Siri as the AI integration layer — where users install and switch between AI providers the way they install apps — that changes the entire competitive landscape. For B2B companies, this means your AI strategy needs to account for Apple's ecosystem as a distribution channel, not just a hardware platform. The companies that build Siri Extensions early will have the same first-mover advantage that early App Store developers had in 2008.

Source: Bloomberg Power On Newsletter, The Verge

Google TurboQuant: 6x Memory Reduction With Zero Accuracy Loss

While the headlines chase chatbot drama, Google Research quietly dropped something that could reshape AI infrastructure economics. TurboQuant, presented at ICLR 2026, is a compression algorithm that reduces AI model memory usage by at least 6x without sacrificing accuracy.

The technical innovation combines two techniques: PolarQuant (which simplifies data geometry through random rotation before compression) and Quantized Johnson-Lindenstrauss (which eliminates residual errors using just 1 additional bit). Together, they solve the key-value cache bottleneck — the memory constraint that limits how much context AI models can process simultaneously.

Why this matters: Memory is the real bottleneck for enterprise AI deployment. Running large models locally requires expensive GPU hardware, and cloud costs scale directly with memory consumption. A 6x reduction means models that required a $40,000 GPU cluster could potentially run on hardware costing a fraction of that. For companies exploring on-premise AI (which you should be, for data sovereignty), TurboQuant-style compression is what makes the economics viable.

Source: Google Research Blog

What This Means for Your Business: 5 Takeaways

1. AI agents are production-ready. Microsoft shipping Cowork through M365 means the "agent" concept has graduated from startup demos to enterprise infrastructure. If you're still in "wait and see" mode, your competitors aren't.

2. Multi-model is the new default. Microsoft using OpenAI for drafting and Anthropic for review. Apple letting users choose their AI provider. The single-vendor AI strategy is dead. Build for model flexibility.

3. Distribution beats capability. Apple doesn't need the best model — they need 2 billion devices. Google doesn't need the flashiest demo — they need the most efficient infrastructure. The AI race is shifting from "who has the best model" to "who has the best platform."

4. Security basics still matter more than AI safety debates. Anthropic's CMS leak is a reminder that your biggest AI risk isn't the model — it's your data pipeline, your access controls, your configuration management. Audit that first.

5. On-premise AI just got more affordable. TurboQuant's 6x memory reduction directly impacts the build-vs-buy calculation. If data sovereignty matters to your industry (it should), the economics of running AI on your own infrastructure are improving fast.

The pace isn't slowing down. If you're a European B2B company trying to figure out which of these developments actually matters for your business — and which ones are just noise — that's exactly what we help with.

Book a 30-minute strategy call and we'll map these trends to your specific situation. No slide decks. No six-month roadmaps. Just clarity on what to do next.