AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
Three things happened this week that any UK board should have on its agenda by Monday. First, the UK Government sent an open letter to business leaders on 15 April warning that AI systems can now find cyber vulnerabilities at “unprecedented speeds”, and told boards to treat AI cyber risk as a governance-level issue. Second, Anthropic disclosed that its Claude Mythos model can independently find and exploit flaws in every major operating system and browser, prompting an emergency meeting between the US Treasury, the Federal Reserve and major bank CEOs. Third, the clock to EU AI Act enforcement is now inside 100 days: the bulk of high-risk obligations become binding on 2 August 2026.
Beyond cyber and compliance, the week was extraordinary on deals and on releases. Amazon lifted its Anthropic commitment to $25 billion in exchange for a $100 billion, ten-year AWS contract, the biggest cloud lock-in in AI history. SpaceX took a $60 billion option on Cursor, rewriting the AI coding market overnight. Tim Cook confirmed his departure as Apple CEO, handing over to John Ternus with an explicit AI mandate. On the tools front, Anthropic shipped Claude Opus 4.7 and launched Claude Design, OpenAI rolled ChatGPT Images 2.0 to every paid plan and gave Codex full desktop-agent powers, and Moonshot released Kimi K2.6, the most capable open-source agentic model to date.
The practical takeaway: elevate AI cyber risk to a board agenda item this week and close your readiness gap against the EU AI Act. Concrete next steps are in “What to Try This Week” below.
What to Try This Week
You do not need a full AI transformation programme to get value from this week’s releases. Three tools stood out in real use. Claude Opus 4.7 is noticeably stronger for complex writing, reasoning and coding tasks. ChatGPT Images 2.0 is materially better at producing usable visuals and handling text in-image. Claude Design is the surprise package: early, but already good enough to test for first-draft slides, one-pagers and visual concepts.
1. Test Claude Opus 4.7 on a real piece of work
Give it something that actually matters: a board memo, a strategy note, a market summary or a messy spreadsheet analysis. The improvement is most obvious on multi-step tasks where weaker models lose structure or nuance. Judge it on output quality, not on benchmarks.
2. Compare ChatGPT Images 2.0 against your current design workflow
Use it to create a LinkedIn graphic, newsletter image, event visual or presentation slide concept. It is not just better at aesthetics; it is also better at following instructions and rendering text cleanly. For many teams, that may be enough to reduce reliance on separate image tools for everyday work.
3. Give Claude Design one hour of serious testing
Ask it to turn a rough idea into a one-pager, sales leave-behind, prototype or presentation concept. It is not replacing a designer for high-stakes brand work, but it is already good enough to accelerate first drafts and reduce blank-page time for commercial teams.
This Week’s Policy & Regulation Brief
UK Government warns boards: elevate AI cyber risk now
On 15 April the UK Government published an open letter to business leaders warning that AI systems are finding vulnerabilities in cyber defences at “unprecedented speeds”, and instructing boards to treat AI cyber threats as a governance-level priority rather than a security-team problem. The letter was co-signed by ministers and echoed by techUK. It is the clearest signal yet that Whitehall expects UK boards to have tested their cyber posture against AI-augmented threats. Treat it as a compliance prompt, not a press release.
EU AI Act: 100 days to enforcement, and Brussels is under siege
The bulk of high-risk obligations under the EU AI Act become enforceable on 2 August 2026, covering risk management, data governance, transparency, human oversight and post-market monitoring. Fines can reach €35 million or 7% of global turnover. In parallel, Amnesty International and civil-society groups sent an open letter on 15 April arguing that the Act’s Omnibus simplification process is gutting core safeguards. Any UK business selling into the EU, or processing EU data, should assume the rules stick as written.
Apple transition: Tim Cook out, John Ternus in, AI the defining job
Apple confirmed on 20 April that Tim Cook will step down as CEO on 1 September. John Ternus, SVP of Hardware Engineering and a 25-year Apple veteran, takes over with the explicit remit of fixing Apple’s AI strategy. AI chief John Giannandrea left earlier in April and his responsibilities have been split across Federighi, Cue and Khan. Apple’s $4 trillion valuation now rests on whether Ternus can close the AI gap. For enterprise IT teams, expect iOS and macOS AI changes to arrive faster than planned.
Amazon locks Anthropic into AWS with $100bn ten-year deal
Amazon announced a fresh $5 billion investment in Anthropic, with up to $20 billion more to follow on commercial milestones, taking total commitment to $33 billion. In return, Anthropic has committed to spend over $100 billion on AWS over the next decade and secured up to 5 gigawatts of Trainium compute. For buyers, this is structural: Anthropic’s infrastructure future is now Amazon’s. It reinforces the case for multi-provider AI procurement and changes the posture between AWS and Microsoft in any enterprise AI deal.
SpaceX takes $60bn option on Cursor, reshapes the coding stack
On 21 April SpaceX announced a partnership with Cursor, the AI coding startup, with an option to acquire it for $60 billion or pay $10 billion for joint development work by year-end. Cursor was valued at $2.5 billion fifteen months ago. OpenAI had reportedly tried and failed to buy the company first. The deal brings Cursor into the Musk empire alongside xAI and Colossus, the 200,000-GPU supercomputer, and will force Microsoft, OpenAI and Google to respond on developer tooling pricing.
Maine passes first US statewide data centre moratorium
Maine’s legislature has approved an 18-month moratorium on new data centres requiring more than 20 megawatts, citing energy grid strain. Governor Janet Mills has a short window to sign or veto. At least 12 other states are weighing similar bills. This is the first serious US planning constraint on AI compute build-out and it lands at exactly the moment Big Tech has committed over $650 billion in AI capex for 2026. Expect the UK grid-capacity debate to intensify in parallel.
Anthropic’s Claude Mythos triggers Treasury-Fed bank summit
Anthropic disclosed that Claude Mythos, currently restricted to a consortium of around 40 partners including JPMorgan, can independently find and exploit vulnerabilities in every major operating system and browser. In testing it identified a 27-year-old OpenBSD flaw. Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened an emergency meeting with US bank CEOs in response. The UK Government’s 15 April letter landed in the same week. If you run critical systems, this is no longer an abstract risk; it is a board reporting line that needs a named owner.
Model & Platform Updates
Anthropic ships Claude Opus 4.7
Released on 16 April, Claude Opus 4.7 is a direct upgrade on 4.6 with stronger software engineering performance, a new “xhigh” reasoning effort setting for complex multi-step work, and a three-times jump in multimodal resolution to 3.75 megapixels per image. A tokeniser change means prompts use slightly fewer tokens for the same content. For teams already building on Claude, it is a drop-in upgrade worth testing this week, particularly if you are running coding agents or processing diagrams, whiteboards, and document scans at scale.
Claude Design: Anthropic steps onto Figma and Canva’s turf
Anthropic launched Claude Design on 17 April as a research preview, a product for creating visual work including prototypes, slides, presentations and one-pagers directly from natural language. It turns a prompt into editable visual output and targets the workflow space currently owned by Figma, Adobe Express and Canva. Early reviews have been strong. For marketing, sales and product teams, this is worth an hour of play this week: if the output is half as good as the demos suggest, it reshapes who needs a designer in the loop.
OpenAI Codex becomes a full desktop agent, but not in the UK
OpenAI’s April Codex update transforms the product into a desktop automation agent. It now operates Mac apps autonomously with its own cursor, remembers preferences, schedules multi-day work, runs parallel agents, and pulls in 90-plus new plugins across Atlassian, GitLab and Microsoft Suite. It also generates images via gpt-image-1.5. Crucially, the desktop-agent feature is not yet available in the EU or UK. That gap is a direct consequence of the AI Act and UK regulatory posture, and worth flagging to anyone on your team running Codex workflows.
ChatGPT Images 2.0 rolls out to every paying plan
OpenAI pushed ChatGPT Images 2.0 to all paid ChatGPT tiers on 21 April, replacing the previous image generator across Plus, Team, Enterprise and Edu. The new model produces materially sharper, more photorealistic output and handles text inside images considerably better than before, which was the most common complaint about the prior version. For anyone using ChatGPT in marketing, internal comms or deck production, this is an immediate free upgrade. Test it against whatever you are currently paying Midjourney or Adobe for.
Moonshot releases Kimi K2.6, the most capable open-source agentic model
Chinese lab Moonshot released Kimi K2.6 on 21 April as a fully open-source, 1-trillion-parameter mixture-of-experts model, supporting 300 parallel sub-agents across 4,000 coordinated steps and 13-hour autonomous coding sessions. It tops the HLE-Full benchmark at 54.0, narrowly ahead of GPT-5.4 and Claude Opus 4.6. For UK businesses that need sovereign or on-premise deployment, this is the most capable open model to date. It matters if your compliance team cannot send data to a US-hosted API.
OpenAI GPT-Rosalind targets life sciences
OpenAI released GPT-Rosalind on 16 April, a frontier reasoning model tuned for life sciences research. Launch partners include Amgen, Moderna, the Allen Institute and Thermo Fisher, with early use cases spanning target identification, protein work and trial design. This follows Novo Nordisk’s partnership announced last week. The pattern is now clear: frontier labs are embedding directly into regulated, high-value verticals rather than selling generic APIs. UK life sciences firms not yet in a similar conversation should expect the competitive gap to widen from here.
Perplexity launches Personal Computer for Mac
Perplexity has launched Personal Computer, a Mac-based extension of its Computer product that can work across local files, native apps and the Comet browser from inside the desktop app, with voice control included. Access is live for Max subscribers, Pro rollout is coming, and Windows is on a waitlist. The bigger shift is strategic: AI tools are moving from answering questions to operating the machine itself.
xAI pushes Grok 4.2 into public beta
xAI has put Grok 4.2 into public beta as an explicitly selectable release candidate, rather than a quiet default upgrade. That matters less for the headline than for the operating model: frontier labs are shipping fast, learning in public and iterating weekly. For enterprise users, that increases optionality, but also raises the bar on version control, testing and governance before deployment.
Cursor and Factory push enterprise coding agents to centre stage
Beyond the SpaceX deal, Cursor’s revenue has gone from $100 million to $2 billion ARR in fourteen months, described as the fastest SaaS ramp in history. Factory AI closed a $150 million round at a $1.5 billion valuation to build enterprise coding agents. The combined message to engineering leaders: autonomous coding agents are now a procurement category, not an experiment. If you have not evaluated them against your current tooling budget in the last quarter, you are already behind your peers.
Quick Hits
Cerebras files for IPO, first major AI-chip listing of the cycle: The Nvidia rival filed its S-1 on 17 April, targeting a mid-May listing with 2025 revenue of $510 million and $24.6 billion in remaining performance obligations.
UK executive sentiment flips negative on AI jobs: Accenture’s Bloomberg-reported poll on 19 April found roughly half of UK business leaders now expect net AI-driven job losses over the decade, up from one in three two years ago.
Meta starts tracking employee keystrokes and mouse movements: Meta confirmed on 21–22 April that it is installing screen and keystroke capture software on all US employee computers to train its agentic AI models, prompting internal backlash and likely regulatory scrutiny.
Adobe ships Firefly AI Assistant across the Creative Cloud: Adobe’s new assistant orchestrates tasks across Photoshop, Premiere and Illustrator from a single prompt, the clearest sign yet that the creative suite is being rebuilt around agent-style workflows.
Deezer reports 44% of new music uploads are AI-generated: Most are flagged as fraudulent, showing how quickly synthetic content is crowding creator platforms and why provenance policies now matter for marketing and brand teams too.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.