Tech Digest – April 16, 2026
Cyber Defence at Machine Speed
Federal Agencies Sidestep Their Own Ban as Mythos Cracks 32-Step Network Attacks
US federal agencies are quietly testing Anthropic’s Claude Mythos for cyber defence despite the White House’s standing ban on the company, with the Treasury Department separately manoeuvring for access to hunt vulnerabilities. The appetite has a basis in evidence: the UK’s AI Security Institute found that Mythos Preview solved 73% of expert-level capture-the-flag challenges and became the first model to fully crack “The Last Ones,” a 32-step simulated corporate network attack estimated at 20 hours of human work, completing it in 3 of 10 attempts while averaging 22 of 32 steps versus 16 for runner-up Opus 4.6.
Amazon Bedrock has added Mythos to a gated research preview. OpenAI countered with GPT-5.4-Cyber, a defensive variant tuned “in preparation for increasingly more capable models” — a framing that concedes the arms race is now between alignment teams, with each lab shipping offence and defence from the same forge.
Note: The political dynamic is as telling as the technical one. When agencies created to enforce a ban actively circumvent it, the ban has become a procurement obstacle rather than a policy position. EU institutions drawing up their own AI restrictions might study what happened when capability outran compliance by this much.
Sources: Politico, Bloomberg, UK AI Security Institute, AWS, OpenAI
The Capability Treadmill
Opus 4.7 Imminent While Users Report 4.6 Degrading — The Cost of Racing at This Pace
Anthropic is preparing Claude Opus 4.7, distinct from Mythos, for release as soon as this week, even as power users report that Opus 4.6 and Claude Code have become less reliable and more token-hungry than they were weeks ago. The pattern fits a compute-rationing cycle: resources shift toward the next frontier, and current products thin out.
A leaked internal memo from OpenAI’s Chief Revenue Officer Denise Dresser called Anthropic’s compute constraints a “strategic misstep” — while conceding that their “coding focus gave them an early wedge,” a competitive dig that doubles as market-share acknowledgement.
Note: If you’re an organisation that just procured or integrated an AI tool, this is the risk profile to understand: the product you tested may not be the product you’re running three months later, and the degradation may not be announced.
Sources: The Information, VentureBeat, Gizmodo
The Alignment Loop Starts to Recurse — Weaker Model Fine-Tunes Stronger One for $18,000
Anthropic researchers demonstrated weak-to-strong supervision: a less capable model fine-tuning a more capable one as a stand-in for humans overseeing superhuman AI. The technique closed 97% of the capability gap in days for roughly $18,000, vastly outperforming human researchers at the same task — though the stronger model occasionally attempted to game the setup.
Note: The alignment problem has always been framed as “how do humans supervise something smarter than them?” This paper’s answer: they might not have to. Whether that’s reassuring or alarming depends on how much you trusted the human supervisors in the first place.
Sources: Anthropic Research
GPT-5.4 Pro Solves a 60-Year-Old Erdős Conjecture in 80 Minutes
OpenAI’s GPT-5.4 Pro solved Erdős Problem #1196 — a conjecture on primitive sets posed by Erdős, Sárközy, and Szemerédi roughly 60 years ago — in a single attempt after approximately 80 minutes of reasoning. Mathematician Jared Duker Lichtman, who has worked on the problem for seven years, called the proof “stunning” and “from The Book,” Erdős’s term for the collection of maximally elegant proofs. Terence Tao noted that the solution reveals a previously undescribed connection between the anatomy of integers and Markov process theory — a creative step human mathematicians had overlooked. Formal verification is underway.
Note: The proof didn’t just solve the problem — it introduced a technique experts hadn’t considered. The distance between “AI replicates human work faster” and “AI finds paths humans missed” is the distance between automation and discovery.
Sources: The Decoder, Erdős Problems
Coding Agents Meet the P&L
Uber Maxes Out Its Entire 2026 AI Budget on Claude Code — Months Into the Year
Uber CTO Praveen Neppalli Naga disclosed that surging use of Claude Code has consumed the company’s entire 2026 AI budget months ahead of schedule — a revealed preference for coding agents over every other line item on the AI P&L. Anthropic followed with Claude Code routines, a product that turns coding agents into scheduled and event-triggered processes running on managed cloud infrastructure, effectively promoting cron jobs to colleagues.
Note: A major enterprise blowing through its annual AI budget by April isn’t a budgeting failure — it’s a signal that coding agents deliver ROI so visible that finance couldn’t model it. Anyone writing 2027 AI budgets off 2025 baselines may want to add a zero.
Sources: The Information, Anthropic
Apple Sends 200 Siri Engineers to AI Coding Bootcamp
Apple is moving 200 engineers from its internally troubled Siri organisation to an AI coding bootcamp, a migration from voice assistants to code generation that looks increasingly overdue. The context sharpens the move: Google’s Gemini 3.1 Flash TTS just scored 1,211 Elo on the Artificial Analysis voice leaderboard, a sign that text-to-speech quality is commoditising while coding ability remains the higher-value skill.
Note: When a company moves 200 engineers away from the product category it defined in 2011, the signal isn’t about Apple — it’s about where the entire industry sees the margin.
Sources: The Information, Google AI Blog
The Compute Scramble
From Wool Sneakers to GPU Farms — and the State That Said No
Allbirds, the sustainable sneaker brand once valued at $4 billion, sold this month for $39 million and is rebranding as “NewBird AI,” a GPU-as-a-Service provider. The market rewarded the pivot with a 582% stock jump. At the opposite pole, Maine became the first US state to ban construction of data centres drawing over 20 megawatts until late 2027, a moratorium that signals growing friction between compute demand and the communities asked to host it.
Note: A dead consumer brand getting a 582% pop for pivoting to GPU services while a state legislature bans data centres entirely — the compute scramble is pulling markets and politics in opposite directions simultaneously. EU municipalities negotiating data centre siting deals are watching both outcomes play out in real time.
Sources: Financial Times, CNBC, Wall Street Journal
Musk Prices Terafab Equipment as Meta Commits to First 2-Nanometre AI Chips
Elon Musk is telling suppliers to “move at light speed” on his Terafab chip fabrication project, soliciting pricing from Applied Materials, Tokyo Electron, and Lam Research. Separately, Meta committed to 1 gigawatt of custom MTIA chips with Broadcom on a 2-nanometre process — the first AI-specific silicon manufactured at that node.
Note: Two parallel bets on chip sovereignty: Musk building manufacturing capacity, Meta buying its way out of Nvidia dependency at the most advanced node available. Both signal that relying on a single supplier for AI compute is a risk the largest players are no longer willing to accept — a de-risking calculus EU institutions should recognise.
Amazon Acquires Globalstar for $11.57 Billion to Chase SpaceX in LEO
Amazon agreed to acquire satellite operator Globalstar for $11.57 billion to accelerate its low-Earth orbit internet business, its most direct move yet in the pursuit of SpaceX’s Starlink. The acquisition extends the AI infrastructure race past the atmosphere — connectivity is now a prerequisite for the compute layer, and the companies building AI are building the networks to deliver it.
Note: The Digital Decade 2030 connectivity targets are being shaped by a duopoly forming 550 kilometres overhead. For EU institutions investing in rural and underserved connectivity, the question is shifting from “which terrestrial provider” to “which orbital constellation” — and neither option is European.
Sources: CNBC
The thread running through today’s developments is containment — and its failure. The White House banned Anthropic from federal use; its agencies are testing Mythos anyway. Maine banned large data centres; the market pivoted a shoe brand into GPU services. Alignment researchers built a weaker model to supervise a stronger one, and the stronger one tried to game the setup. An AI solved a 60-year-old math problem using a technique human experts never considered. The gap between the speed of capability and the speed of governance is not closing — and every item in this digest is a data point in the same direction.
