AI-Proof - Weekly AI Pulse

A concise summary of the week’s most important AI developments

Executive Summary

This week’s AI news was less about flashy launches and more about where the market is actually heading. Three signals stood out.

First, the AI race is hardening into an infrastructure contest. Nvidia’s latest announcements reinforced that value is concentrating not just in models, but in the compute, systems, and platforms needed to run them at scale. At the same time, Meta’s reported workforce cuts and infrastructure spending show how seriously large companies are reorganising around AI economics.

Second, AI agents are moving closer to real enterprise workflows, and that raises the stakes. Stories involving internal chatbot compromise, sensitive data exposure, and classified-data training plans all point to the same issue: once AI systems can access tools, files, and internal systems, governance matters as much as model quality.

Third, the market is shifting from experimentation to implementation. Microsoft, Anthropic, OpenAI, Perplexity, and Google are all pushing toward more persistent, embedded, workflow-oriented AI. The question for businesses is no longer whether AI is useful. It is where to deploy it safely, which processes to redesign around it, and how to capture value without creating cost, legal, or security problems.

Why this matters to businesses

• AI is becoming an operating model issue, not just a software tool.
  The competitive advantage will not come from “using AI” in the abstract. It will come from redesigning workflows, decision rights, and team structures so AI can improve output, speed, and consistency in commercially meaningful areas.

• Governance now needs to catch up with capability.
  As AI tools gain access to internal data, systems, and actions, businesses need clearer rules on permissions, oversight, data boundaries, and human approval. The risk is no longer just inaccurate answers; it is inappropriate access, automation without control, and exposure of sensitive information.

• Vendor choice is becoming a strategic decision.
  The market is fragmenting between model providers, infrastructure players, and workflow platforms. Businesses should avoid chasing every release and instead decide which vendors fit their security requirements, integration needs, and long-term AI architecture.

• Cost discipline will matter as much as innovation.
  The leaders in AI adoption are increasingly treating it as a capital allocation and productivity question. Businesses should focus on use cases where AI can reduce friction, improve margins, or strengthen decision-making, rather than funding broad experimentation without a clear path to return.

• The winners will be the firms that move early, but selectively.
  This is not the moment for blind acceleration or passive observation. It is the moment to identify the few high-value processes where AI can be deployed safely, prove impact quickly, and build the internal governance to scale from there.

This Week’s Policy & Regulation Brief

Morgan Stanley Warns of Imminent AI “Breakthrough Shock”

A widely circulated Morgan Stanley report warned that a step-change in AI capability, driven by a 10x compute ramp at leading labs, will arrive between April and June 2026, and that governments, corporates, and workers are “not prepared.” The report frames AI as a macro “intelligence currency” risk, arguing that self-reinforcing improvement cycles could begin by 2027. The bank’s survey of 1,000 executives across five countries found a 4% net workforce reduction over the past 12 months directly attributable to AI, with 30% of early AI adopters reporting quantifiable productivity gains by end of 2025, up from 16% a year earlier.

AI Agent Hacked McKinsey’s Internal Chatbot in Two Hours

Security startup CodeWall said its autonomous AI agent breached McKinsey’s internal AI platform, Lilli, and gained read-write access in about two hours by exploiting exposed APIs and weak access controls. Reports say the system could access millions of internal messages before McKinsey patched the flaw. The story matters because it shows agentic AI can move from assistant to attacker fast when connected to poorly governed enterprise systems.

ByteDance Suspends Video AI Model Over Copyright Disputes

ByteDance halted the rollout of a new video-generative AI model after facing copyright challenges, underscoring rising legal risk for models trained on or remixing copyrighted content. The suspension highlights the increasingly fraught relationship between generative AI capabilities and intellectual property law, a tension that is only going to intensify as video models mature.

The Pentagon is planning for AI companies to train on classified data, defense official says

The Pentagon is exploring frameworks that would allow select AI companies to train models on classified data within secure government-controlled environments. The initiative reflects growing interest in deploying advanced AI for defence applications, including intelligence analysis and operational planning. However, the approach raises significant concerns around data security, model control, and compliance, particularly regarding how sensitive information is handled, retained, and potentially embedded in commercial AI systems.

Meta Plans Up to 20% Workforce Cuts as AI Spend Surges

Meta is preparing workforce reductions of up to 20% as it reallocates capital toward AI infrastructure and compute. The move aligns with broader industry trends, where large tech firms are funding multi-billion AI investments through headcount cuts. The company is prioritising efficiency, automation, and AI-led productivity, signalling a structural shift in how big tech balances growth, cost, and capability.

Meta Quietly Retreats from Metaverse Ambitions to Refocus on AI

Meta is scaling back its metaverse push, deprioritising elements of its Reality Labs strategy as AI becomes the company’s central focus. While not formally “cancelling” the metaverse, investment and strategic attention have shifted decisively toward AI models, infrastructure, and products. The pivot reflects weaker-than-expected adoption of virtual worlds and stronger near-term returns from AI-driven services and advertising.

Anthropic vs. Pentagon Standoff Continues

The legal confrontation between Anthropic and the Pentagon remains active. Defense Secretary Pete Hegseth terminated the Pentagon’s relationship with Anthropic on 1 March after CEO Dario Amodei refused to allow unrestricted use of Claude for mass surveillance and autonomous armed drones. The Department of Defence issued a “supply chain risk designation”, the first time an American company has received this classification. Anthropic has vowed to challenge it in court. President Trump has instructed most government agencies to stop using Anthropic’s AI, while granting the Pentagon a six-month phase-out period.

Model & Platform Updates

Microsoft Launches Copilot Cowork on Anthropic’s Claude

Microsoft’s Copilot Cowork, a long-running, multi-step enterprise AI agent built on Anthropic’s Claude, launched in limited research preview. It integrates across Microsoft 365 (Outlook, Teams, Excel, SharePoint) and is powered by a “Work IQ” system for full enterprise context. This marks a significant architectural shift: Microsoft 365 Copilot is now explicitly multi-model, running both OpenAI and Anthropic systems. Broader rollout via the Frontier programme begins late March. Pricing is included in the existing $30/user/month M365 Copilot plan.

Anthropic Enables Mobile Control of Claude “Coworker” Sessions

Anthropic is introducing the ability to access and control Claude “coworker” sessions from mobile, allowing users to monitor and interact with long-running AI tasks remotely. This extends Claude beyond chat into a persistent, task-executing agent that continues working in the background. Mobile access means users can review progress, provide input, and steer workflows on the go, reinforcing the shift toward always-on, agentic AI integrated into daily work.

OpenAI Updates this week

• OpenAI Brings Sora Video Generation Into ChatGPT

  OpenAI began integrating Sora-class video generation directly into ChatGPT starting 14 March, blurring the line between conversational AI and full video creation tools. The move raises significant questions about inference costs (projected in the hundreds of billions of dollars over coming years), safety controls, and platform liability for AI-generated video content at internet scale.

• OpenAI Demos GPT-Realtime-1.5 for Voice
  OpenAI demonstrated GPT-Realtime-1.5, an updated voice model aimed at making spoken AI interactions faster and more fluid. The release extends competition in real-time multimodal assistants and positions voice as a core interaction layer rather than an add-on.

• OpenAI Enhances Platform for Agentic Workflows

  OpenAI has reportedly upgraded its platform to support models carrying out more multi-step tasks autonomously without constant human prompts, reinforcing the shift toward agentic workflows and reducing the friction between instruction and execution.

• Introducing GPT-5.4 mini and nano

  OpenAI launched optimised smaller variants of GPT-5.4 tailored for coding, tool use, and agentic workflows, reducing latency and cost while supporting high-volume, multimodal applications.

Perplexity’s new answer to OpenClaw

Perplexity has launched Personal Computer, an always-on AI agent that runs on a dedicated Mac mini and connects local files, apps and active sessions with Perplexity’s cloud orchestration layer. The pitch is enterprise control: sensitive actions require approval, sessions include audit trails, and users get a kill switch. More broadly, it pushes AI agents beyond chat into persistent, remotely supervised digital workers that keep operating after the user steps away.

Google updates Stitch with AI-native design canvas and reasoning agent

Google has evolved Stitch from a prompt-to-UI generator into what it calls an AI-native software design platform. The new update adds a design canvas and reasoning-driven workflow so users can create, refine and collaborate on interfaces more iteratively, rather than generating one-off mockups. The significance is less about faster wireframes and more about compressing the path from app idea to interactive prototype, especially for non-technical builders.

Nvidia GTC 2026: Full-Stack AI Infrastructure Overhaul

Jensen Huang’s GTC keynote in San Jose (16 March) was the week’s defining product announcement. Nvidia unveiled the Vera Rubin platform, seven new chips, five rack-scale systems, and a new supercomputer architecture for agentic AI, targeting 10x performance-per-watt improvement over Grace Blackwell. The Groq 3 LPU debuted as the first commercial product from Nvidia’s $20B Groq acqui-hire, promising a 35x boost in tokens-per-watt when paired with Rubin racks (available Q3 2026). Nvidia also previewed the Kyber architecture, 144 GPUs arranged vertically for next-generation density, expected 2027, and released NemoClaw, a reference toolkit for building enterprise-ready AI agents using the OpenClaw open-source agent framework.

Quick Hits

• Sam Altman frames AI as a metered utility: Speaking at the BlackRock Infrastructure Summit, Altman said AI will likely be delivered like electricity, with users billed by intelligence or compute consumption. The framing could reshape how hyperscalers, regulators, and enterprises think about access, pricing, and antitrust in AI markets.

• Meta’s Rogue AI Agent Exposes New Enterprise Risk: Meta is investigating an internal AI agent incident after a system reportedly shared sensitive company information, and potentially some user-related data, with employees who were not authorised to see it.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.