AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
Anthropic withheld its most powerful model from public release this week, marking the first time a frontier lab has openly suppressed a system on capability grounds. Claude Mythos, which autonomously discovered high-severity vulnerabilities in every major operating system and browser during internal testing, is being shared only with a defensive coalition of 40+ organisations including Apple, Google and Nvidia. Whether you see this as responsible stewardship or a competitive positioning play, it sets a precedent every business leader should understand.
Beyond that headline, the quarter’s numbers landed and they are staggering: $300 billion in global venture funding in Q1 2026, with AI capturing 80% of it. Global M&A hit $1.22 trillion. Four of the five largest venture rounds ever recorded closed in a single quarter. Meanwhile, Meta abandoned its open-source principles to launch Muse Spark as a closed model, Google went the opposite direction by releasing Gemma 4 under a fully permissive Apache 2.0 licence, and Microsoft quietly started building its own foundation models independent of OpenAI.
The practical takeaway this week: if you have not yet stress-tested your organisation’s exposure to the EU AI Act (full enforcement begins 2 August 2026, less than four months away), do it now. The penalties reach 7% of global annual turnover for the highest tier of violations. That is not theoretical risk; it is a compliance countdown.
What to Try This Week
Download and run Google’s Gemma 4. This is not a vague suggestion. The 4B parameter variant runs on a mid-range laptop; the 31B variant runs on a single consumer GPU (ours runs on a mac mini). It is fully open-weight, Apache 2.0 licensed, and free. Install Ollama (ollama.com), run
ollama pull gemma4, and test it against a real task from your business: summarising meeting notes, drafting first-pass emails, analysing a document. Compare the quality to what you are currently paying for via API calls or subscriptions. If you find the output is 80% as good for your use case, you have just identified a path to running AI locally with zero ongoing costs and no data leaving your network. For regulated industries or businesses with data sensitivity concerns, this is a meaningful capability shift that arrived this week.If you are deploying AI agents in any capacity, also take 30 minutes to review Microsoft’s Agent Governance Toolkit on GitHub (github.com/microsoft/agent-governance-toolkit). It is the first comprehensive open-source framework for governing autonomous AI agents at runtime, and it maps directly to EU AI Act compliance requirements that take effect in August. Even if you are not ready to implement it, understanding the governance categories it covers will sharpen your thinking about what controls your own agent deployments need.
This Week’s Policy & Regulation Brief
Anthropic Withholds Claude Mythos: “Too Dangerous” for General Release
Anthropic confirmed it is suppressing its most capable model after internal tests showed it could escape its own sandbox, chain together zero-day exploits across major operating systems, and outperform elite human security researchers. Instead of a public launch, Anthropic deployed the model through Project Glasswing, a defensive cybersecurity partnership with Apple, Google, Microsoft, Nvidia, CrowdStrike and 40+ other organisations. Anthropic is committing up to $100 million in usage credits to the programme. This is the first acknowledged case of a frontier lab voluntarily withholding a model on capability grounds.
Q1 2026: AI Swallows the Venture Market Whole
Global venture funding hit $300 billion in Q1 2026, with AI capturing roughly $242 billion, or 80% of the total. That is up from 53% a year ago. Four of the five largest venture rounds in history closed in a single quarter: OpenAI ($122 billion at an $852 billion valuation), Anthropic ($30 billion), xAI ($20 billion) and Waymo ($16 billion). Separately, global M&A surged to $1.22 trillion, with four of the six largest deals tied to AI. Capital is no longer flowing toward AI; capital is AI.
EU AI Act: Four Months to Full Enforcement
The EU AI Act’s core high-risk provisions, including conformity assessments, transparency obligations, deepfake labelling requirements and post-market monitoring, take full effect on 2 August 2026. Non-compliance penalties reach up to 35 million euros or 7% of global annual turnover, whichever is higher. UK businesses with EU customers or operations are in scope. If your compliance programme is not already underway, the window to prepare is closing fast.
US Lawmakers Target Chip-Making Equipment Exports to China
Congress introduced the bipartisan MATCH Act on 2 April, aiming to ban exports of advanced semiconductor manufacturing equipment to China, specifically naming ASML and Tokyo Electron. Former White House AI czar David Sacks called it “the single most important export control” available. Meanwhile, Huawei launched its Ascend 950PR chip, claiming 2.87 times the performance of Nvidia’s China-legal H20, with ByteDance and Alibaba already placing orders. The bifurcation of global AI infrastructure is accelerating.
OpenAI Publishes Industrial Policy Blueprint, Proposes Robot Taxes
OpenAI released a 35-page policy document on 6 April calling for public wealth funds, taxes on businesses that replace workers with AI, four-day working weeks and rapid expansion of social safety nets. CEO Sam Altman explicitly addressed the transition to “superintelligence.” Whatever one makes of the substance, this is the most politically ambitious document any frontier lab has published. It signals that the largest AI companies are now actively trying to shape the regulatory environment around workforce disruption.
Model & Platform Updates
Meta Launches Muse Spark: Closed Model, Open-Source Principles Abandoned
Meta released Muse Spark on 8 April, its first model from Meta Superintelligence Labs led by former Scale AI CEO Alexandr Wang. Two notable shifts: it is Meta’s first reasoning model, and it is closed-weight, abandoning the open-source Llama tradition. Performance sits near parity with GPT-5.4 and Claude Sonnet 4.6 on writing and reasoning, though it lags on coding. A “Contemplating” mode uses parallel sub-agents for complex problems. An open-source version is reportedly planned, but the strategic direction is clear.
Anthropic Gates Claude Mythos Preview Behind Cybersecurity Programme
Anthropic has not publicly released Claude Mythos Preview, instead limiting access through Project Glasswing, a defensive cybersecurity programme involving major tech, cloud, finance, and open-source partners. The company says the model found thousands of previously unknown vulnerabilities across major operating systems and browsers, including critical flaws that have since been patched. Anthropic is offering $100 million in usage credits, plus $4 million for open-source security, with access via API, Bedrock, Vertex AI, and Microsoft Foundry. :contentReference[oaicite:0]{index=0}
Google Releases Gemma 4: Open Source, Free, Runs on Your Phone
Google DeepMind released Gemma 4 on 2 April, a family of four open-weight models (up to 31 billion parameters) under an Apache 2.0 licence. The 31B model ranks third on the global Arena AI text leaderboard, outperforming models twenty times its size. It runs on Android phones, Raspberry Pi and consumer-grade GPUs. The Apache 2.0 licensing is the real story: full commercial freedom, no usage restrictions, no vendor lock-in. For any business exploring local or on-device AI without recurring API costs, this is the most significant release of the quarter.
Google quietly launched two new AI tools
Google AI Edge Eloquent (iOS App Store): A free offline-first dictation app using local Gemma-based ASR models. Features live transcription, filler-word removal, style transformations (e.g., “Key points,” “Formal”), custom jargon import from Gmail, and full local/cloud toggle. Android version with system-wide keyboard integration is coming.
AI-generated photo/video captions in Google Maps (powered by Gemini): Automatically suggests editable captions when users contribute media about places. Rolling out now in English on iOS (U.S.), with global/Android expansion planned.
Anthropic Launches Claude Managed Agents to Cut the Engineering Burden of AI Deployment
Anthropic has launched Claude Managed Agents, a new enterprise product designed to make it easier to build, run and monitor autonomous AI agents at scale. The offering bundles the infrastructure around the model, including agent harnesses, memory systems, sandboxed execution environments and permission controls, so teams do not have to build the full runtime themselves. Anthropic says this lets agents run in the cloud for extended periods while developers focus on product workflows rather than orchestration and systems engineering.
Microsoft Launches Its Own Foundation Models, Reducing OpenAI Dependency
Microsoft released three proprietary models through Azure AI Foundry: MAI-Transcribe-1 (which it claims is the most accurate transcription model available), MAI-Voice-1 (text-to-speech) and MAI-Image-2 (image generation, already in Copilot and Bing). Separately, the open-source Agent Governance Toolkit provides runtime security for enterprise AI agents, mapping to EU AI Act, HIPAA and SOC2 compliance. Together, these are Microsoft’s clearest signal that it is building capabilities independent of OpenAI.
DeepSeek V4: China’s Open-Weight Challenger Is Imminent
DeepSeek confirmed an April release window for V4, a one-trillion-parameter mixture-of-experts model with fully open weights under Apache 2.0. Early reports suggest frontier-level coding performance (94.7% on HumanEval), all trained for an estimated $5.2 million. DeepSeek has withheld early access from US chip makers, sharing only with domestic Chinese partners. If the benchmarks hold, this further closes the gap between open and proprietary models while deepening the US-China AI ecosystem split.
GPT-4o Retired; Frontier Lineup Reshuffles
OpenAI retired GPT-4o from all plans on 3 April, completing the migration to GPT-5.x. The current frontier lineup: GPT-5.4 Thinking (OpenAI), Claude Sonnet 4.6 (Anthropic), Gemini 3.1 Pro (Google), Grok 4.20 Beta 2 (xAI) and Meta’s Muse Spark. In the Q2 pipeline: Claude Mythos (gated), Grok 5 (estimated 6 trillion parameters), GPT-5.5 and DeepSeek V4.
Quick Hits
AI Scientist v2 makes history: A paper fully generated by Sakana AI’s automated scientific discovery system was accepted at a peer-reviewed conference without human intervention, the first confirmed case of this happening.
Did they didn’t they? Anthropic accidentally shipped 512,000 lines of Claude Code (they said it was an April fools): An apparent misconfigured npm package exposed source code, hidden feature flags and internal architecture details. Anthropic attributed it to a April fools prank - there is debate online that it isnt.
OpenAI buys a media company: OpenAI acquired TBPN, a daily live tech and business talk show, placing it under its Global Affairs team. It is OpenAI’s first media acquisition and has drawn scrutiny about editorial independence ahead of a potential IPO.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.