Tech Digest – May 1, 2026
Frontier AI Enters the National Security Stack
GPT-5.5 Matches Mythos on Cybersecurity — NSA Already Hunting Vulnerabilities in Microsoft Software
The UK’s AI Security Institute evaluated an early checkpoint of OpenAI’s GPT-5.5 and found it matched Anthropic’s unreleased Mythos on advanced cybersecurity tasks, achieving a 71.4% success rate at Expert difficulty compared with Mythos’s 68.6%. GPT-5.5 is only the second model to complete a 32-step corporate network attack simulation end-to-end — a chain that would take a human expert roughly 20 hours. Separately, the NSA has begun testing Mythos itself to find vulnerabilities in Microsoft software and other widely used programs under what CISA briefings describe as autonomous red-teaming at scale. Anthropic has kept Mythos out of public hands after concluding that attackers could exploit it, and the White House is opposing a plan to extend access to roughly 70 additional organisations.
Note: The AISI found a universal jailbreak for GPT-5.5’s cyber capabilities that took just six hours to develop. Offensive cyber capability is now emerging as a side effect of general improvements in reasoning and coding — which means every frontier model upgrade is also a cyber-capability upgrade, whether the lab intended it or not.
Sources: UK AI Security Institute, Bloomberg
Chip Supply Chain Splits Along National Lines
Huawei Takes China’s AI Chip Crown as Intel Posts Its Best Month in 55 Years
Huawei is set to capture the largest share of China’s AI chip market this year, targeting roughly $12 billion in AI chip revenue as sales jump 60%. Nvidia’s share of the Chinese market has fallen from 95% to approximately 55% after Washington blocked H20 sales and Nvidia has struggled to find a compliant replacement. On the other side of the bifurcation, Intel posted a 114% share price gain in April — the best month in its 55-year history — pushing its market cap past $470 billion. Sandisk, supplying the memory layer underneath, reported quarterly revenue up 251% year over year. Even the largest AI lab cannot escape the squeeze: Demis Hassabis acknowledged that Google lacks sufficient TPUs to run two frontier model families simultaneously, explaining why Gemma stays compact while Gemini absorbs the silicon.
Note: The supply chain is now splitting faster than procurement frameworks can adjust. Nvidia’s China share halving in under two years shows how quickly export controls can restructure a market. For any institution specifying AI hardware requirements, “Nvidia” is no longer a universal answer — it is a geopolitical position.
Sources: Financial Times, CNBC, MarketWatch
Physical Automation at Every Scale
1X Opens Humanoid Factory Targeting 100,000 Units by 2027 — SoftBank Builds Robots to Build Data Centres
1X Technologies opened a 5,400 m² factory in Hayward, California, with a target of 10,000 home humanoid robots this year and 100,000 by the end of 2027. Consumer shipments are expected before the holidays. Separately, SoftBank is assembling Roze AI, a new venture that will deploy autonomous robots to construct data centres, and is already eyeing a $100 billion IPO before a single robot has shipped — the AI capital expenditure cycle now literally building itself. Meta, meanwhile, sold another $25 billion in bonds to fund AI infrastructure, a reminder that the demand these robots would serve shows no sign of plateauing.
Note: If 1X’s timeline holds, 100,000 humanoids enter the consumer market by 2027. Even a fraction reaching the care or domestic services sector would shift labour shortage assumptions in areas where EU workforce gaps are most acute — eldercare, facility maintenance, logistics support.
Sources: Bloomberg, TechCrunch, Bloomberg
First Electric Air Taxi Completes JFK-to-Manhattan Flight in 15 Minutes
A Joby Aviation eVTOL prototype completed the first electric air taxi flight from JFK Airport, touching down at Manhattan’s West 30th Street heliport 15 minutes later. The flight places operational urban air mobility testing inside one of the world’s most complex and heavily regulated airspace environments.
Note: EASA’s U-space regulatory framework for urban air mobility is still in early deployment across the EU. If eVTOL operations prove viable in JFK airspace — arguably the hardest regulatory environment to crack — the proof point accelerates pressure on EU certification timelines.
Sources: Flying Magazine
Clinical AI Crosses the Diagnostic Threshold
DeepMind Launches AI Co-Clinician as Harvard Study Finds AI Outperforms Physicians
Google DeepMind launched an AI co-clinician designed to function as a collaborative member of the clinical care team under expert supervision — not a standalone tool, but an integrated participant in patient management. The deployment arrives alongside a Harvard and Beth Israel Deaconess Medical Center study finding that OpenAI’s o1 series outperformed both human physicians and older AI models across diagnosis and management on real clinical cases.
Note: The framing choice matters. DeepMind positioned this as a team member, not a replacement — a design decision aimed at earning clinician trust and regulatory clearance. Institutions deploying clinical AI face the same choice: standalone diagnostic oracle or integrated colleague. The evidence increasingly favours the model that makes doctors better rather than the one that tries to replace them.
Sources: Google DeepMind, NPR
Training Practices Under Oath
Musk Admits xAI Distilled OpenAI’s Models to Train Grok
During testimony in the ongoing OpenAI trial, Elon Musk acknowledged under oath that xAI used distillation of OpenAI’s models as part of training Grok, calling it a general practice among AI companies. Distillation — using a larger model’s outputs to generate synthetic training data for another — is common but increasingly contested. In the same proceedings, Judge Yvonne Gonzalez Rogers told Musk’s lawyers the court would not entertain arguments about AI catastrophe or extinction.
Note: Distillation sits in a legal grey zone that the EU AI Act’s transparency and documentation requirements will eventually force into the open. For institutions evaluating AI vendors, the question is sharpening: can your vendor document the provenance of its training process? If a frontier model was partly trained on another company’s outputs, what does that mean for IP liability downstream?
Sources: Wired, New York Times
Today’s threads share a common shift: from capability to consequence. Frontier models are hunting real vulnerabilities in production software, not benchmarks. The chip supply chain is splitting along lines that map to diplomatic alliances, not cost optimisation. Robots are building the infrastructure that builds more robots. And in a courtroom, the question is not whether distillation works — it is whether it was authorised. The technology questions are being answered, one after another. The institutional ones are arriving faster than anyone is writing policy for them.
