Tech Digest – May 14, 2026
Capability Without Ceiling
UK Safety Institute Confirms AI Capability Doubles Every Four to Five Months — Mythos Clears Both Cyber Ranges
The UK’s AI Security Institute published its latest capability evaluation, finding that frontier AI models’ autonomous cyber capability has been doubling roughly every four to five months since reasoning models emerged in late 2024. Anthropic’s Mythos Preview became the first model to complete both of AISI’s multi-step cyber-attack simulations end-to-end — exercises the institute estimates would take a skilled human around 20 hours each. GPT-5.5 cleared one of the two. The institute found no clear upper bound on either model’s performance; the limiting factor appears to be token budget, not ability. Nvidia, whose chips underpin most of this capability, crossed a $5.5 trillion market cap on the same day.
Note: The distinction between “limited by tokens” and “limited by ability” matters for institutional planning. If the ceiling is economic rather than technical, it moves with every price drop. Threat models built around what AI cannot do may need rebuilding around what it cannot yet afford to do.
Sources: UK AI Security Institute, Ethan Mollick
Recursive Superintelligence Emerges From Stealth With $650M — Former Lab Leads Bet on Self-Improving AI
Recursive Superintelligence launched with $650 million in Series A funding at a $4.65 billion valuation, backed by Alphabet’s GV, Greycroft, Nvidia, and AMD Ventures. Founded by former Salesforce Chief Scientist Richard Socher alongside alumni from OpenAI, Google DeepMind, Meta AI, and Uber AI, the company operates from London and San Francisco. Its stated mission: building AI systems that conduct experiments on how to safely improve themselves — the explicit pursuit of recursive self-improvement as the fastest path to artificial superintelligence.
Note: The London office matters. Europe’s largest AI investments tend to arrive as branch offices of US firms. A company explicitly pursuing superintelligence with a London research base invites a question EU policymakers have not yet answered: does the AI Act’s risk framework extend to systems designed to recursively improve beyond their initial assessment?
Sources: SiliconANGLE, CNBC, TechFundingNews
Agentic Products at Scale
Anthropic Launches Claude for Small Business — AI Agents Now Handle Payroll, Invoicing, and Campaigns
Anthropic released Claude for Small Business, a single-toggle integration that connects Claude to QuickBooks, PayPal, HubSpot, Canva, DocuSign, and the Google and Microsoft productivity suites. The product can run payroll, close books, chase invoices, and execute sales campaigns — operational tasks that typically require multiple employees or outsourced services. Separately, the Wall Street Journal reported that Anthropic has overtaken OpenAI inside Ramp’s enterprise customer base, 34.4% to 32.3%, suggesting enterprise AI spending is shifting in real time.
Note: This is not another chatbot integration — it is a pre-wired operations layer. A sole proprietor who previously needed a bookkeeper, a marketing contractor, and an admin assistant can now route those functions through a single AI agent with access to production systems. For any institution that advises or funds small businesses, the support model assumptions are changing.
Sources: Anthropic, Wall Street Journal
Robotics Crosses the Deployment Line
Robot App Store, 8-Hour Factory Shifts, and a Fully Automated Lab — Three Deployment Milestones in One Day
Unitree opened UniStore, the world’s first robot task-and-motion app store, with 24 downloadable behaviours for its humanoid and quadruped units — owners install capabilities the way smartphone users install apps, with a stated goal of 100,000 connected devices within three years. Figure AI live-streamed a team of Helix-02 humanoids running a full autonomous 8-hour package-sorting shift, with 300,000 concurrent viewers watching robots match human pace at roughly three seconds per package — all inference running onboard, no cloud connection. And the Institute of Science Tokyo opened the world’s first fully automated medicine lab at its Yushima campus, where 10 humanoid robots led by the Maholo LabDroid run 1,000 experiments daily with no human staff inside the facility.
Note: Three thresholds crossed simultaneously: robots as a software platform (install a skill, not a machine), robots as shift workers (8 hours, zero intervention), and robots as researchers (cell culture, reagent dosing, around the clock). The question for institutions is no longer whether robots can work — it is how fast deployment scales once the platform economics click.
Sources: Pandaily, TechTimes, Interesting Engineering, Ed Ludlow (Bloomberg)
Digital Sovereignty & AI Governance
Poland Advances 3% Digital Services Tax on US Tech Giants, Defying Washington
Poland’s Finance Minister Andrzej Domanski confirmed the government will proceed with legislation imposing up to 3% tax on digital platforms that sell advertising, process user data, or enable online sales. The tax targets companies with global digital revenues exceeding €1 billion and Poland-sourced revenues above approximately €6.9 million, and is projected to raise around €400 million annually starting in 2027. Poland is pressing ahead despite explicit US threats of retaliation.
Note: This is the first EU member state to advance a standalone digital services tax since the OECD Pillar One negotiations stalled. If the legislation survives US pressure, it becomes a template. Other member states watching the outcome are running a low-risk observation on whether digital sovereignty through taxation is enforceable within the current transatlantic dynamic.
Sources: Bloomberg
OpenAI Proposes IAEA-Style AI Governance Body as Nvidia’s CEO Joins China Delegation
OpenAI’s head of global affairs Chris Lehane proposed a global AI governance body modelled on the International Atomic Energy Agency — US-led but including China — while Nvidia CEO Jensen Huang joined the US president’s delegation for a direct meeting with Xi Jinping. The two moves landed the same week that a quarter of Washington’s 13,000 federal lobbyists were found to be working AI issues, up from 11% in 2023, according to the New York Times.
Note: The IAEA comparison reveals the framing: AI as a dual-use technology requiring international inspection regimes. But the IAEA works because enrichment facilities are physical and countable. AI capability is neither. The proposal may matter less for its design than for who made it — the company that stands to be most constrained by such a body is the one requesting it be built.
Sources: Bloomberg, CNBC, New York Times
Workforce Signals
Meta Employees Organise Protest Against Mouse-Tracking Software That Trains Their Replacement
Meta’s US employees are organising against internal mouse-tracking software that captures every cursor movement on their work machines. The data is being used to train AI models — effectively drafting employees’ daily workflows into the system being built to automate those same roles. The protest represents one of the first organised employee actions targeting the specific mechanism of AI-driven displacement: not layoffs announced from above, but behavioural data harvested from below.
Note: Most workforce displacement assumes a top-down decision: management replaces roles with AI. This inverts it. The replacement is being trained in real time on the people it will replace, using data they generate involuntarily. It surfaces a question labour law has not caught up with: who owns the pattern of how you move a mouse?
Sources: Reuters
The measuring has started. The UK government measured capability doubling and found no ceiling. Figure measured a full shift and streamed it to 300,000 people. Poland measured the tax gap and wrote legislation. Meta employees measured the surveillance and organised. The era of speculation about AI’s institutional impact is giving way to an era of measurement — and measurements demand responses. When you can time the doubling, audit the shift, and price the tax, the question is no longer whether to act.
