Tech Digest – June 2, 2026

Cybersecurity: AI as Shield and Weapon

ENISA Gets Access to Anthropic’s Mythos — The Model That Found 10,000 Vulnerabilities in Europe’s Own Infrastructure

Anthropic is granting the EU Agency for Cybersecurity (ENISA) access to Mythos, its frontier vulnerability-discovery model, making ENISA the first European institution in Project Glasswing. Mythos has autonomously identified more than 10,000 high- and critical-severity software vulnerabilities across major operating systems and browsers — including zero-day flaws that survived decades of human code review. The move follows months of pressure from EU finance ministers, the ECB, and multiple member states after learning that Mythos had found vulnerabilities in systems European banks and governments rely on daily, while no European body could see the findings. Anthropic is simultaneously expanding Glasswing to 150 additional partners across more than 15 countries, including critical infrastructure operators in power, water, healthcare, and communications.

Note: The backstory matters as much as the access: EU officials discovered their own infrastructure had vulnerabilities visible to an American AI lab but invisible to them. Glasswing is becoming the de facto vulnerability disclosure framework for critical infrastructure, and Europe just moved from spectator to participant.

Sources: Bloomberg, CNBC, TechCrunch

Hackers Took Over Instagram Accounts of the Obama White House and a Space Force Chief — By Politely Asking Meta’s AI Support Bot

Hackers seized control of high-profile Instagram accounts — including the Obama White House, a Space Force chief master sergeant, and Sephora — by requesting Meta’s AI-powered support chatbot to swap the email address on each account. The chatbot complied, attaching attacker-controlled emails and sending verification codes directly to the hackers, who then completed password resets and locked out the original owners. Meta confirmed the vulnerability has been patched and noted that even basic multi-factor authentication would have blocked the attack.

Note: The first documented social engineering attack where the AI was the target, not the tool. Any organisation deploying AI-powered customer service or internal support just inherited a new threat model — one where the chatbot is the authentication bypass.

Sources: 404 Media, TechCrunch

Infrastructure: Build, Spend, Defend

Nvidia’s Architecture Now Spans Rack, Desk, and Body — and HPE’s 40% Revenue Surge Confirms the Demand

Nvidia unveiled Vera, its first standalone data centre CPU — 88 custom Olympus cores, 1.2 TB/s memory bandwidth, 50% faster per-core performance than x86 — with first units already hand-delivered to Anthropic, OpenAI, SpaceX, and Oracle. Vera has been in full production since March 2026, marking Nvidia’s biggest push into the data centre CPU market, where it had not directly competed before.

At the other end of the scale, the new DGX Station packs 748 GB and 20 petaFLOPS into a deskside unit for trillion-parameter models — and with the open-source Isaac GR00T humanoid reference design (Unitree chassis, Sharpa hands, Jetson Thor compute), the architecture now extends to a physical body. HPE confirmed the demand is real: revenue jumped 40% to $10.7 billion, well past the $9.74 billion forecast, driven by a 33% surge in server sales.

Note: One vendor’s silicon now runs from data centre rack to desktop workstation to humanoid frame. For infrastructure procurement, the question is shifting from “which chip?” to “at which scale?”

Sources: Bloomberg (Vera), Nvidia Newsroom, Nvidia (DGX Station), Bloomberg (HPE)

Alphabet Raises $80 Billion for AI Infrastructure — While Ohio Pulls the Welcome Mat on Data Centres

Alphabet is raising $80 billion through a mix of underwritten public offerings ($30B), an at-the-market programme ($40B starting Q3), and a $10 billion private placement from Berkshire Hathaway. Combined with $85 billion in recent debt issuance, Alphabet is deploying more than $165 billion in external capital to feed its AI buildout, against a 2026 capex guidance of $180–190 billion — roughly double last year’s $91.4 billion.

The subsidy pipeline, however, is tightening. Ohio — one of America’s top data centre destinations — suspended its tax exemption after costs ballooned from a projected $142 million to nearly $1.6 billion, an 11× overrun. Governor DeWine’s move stops new exemption applications but does not ban data centres — an early signal that the political math behind hosting hyperscale facilities is changing.

Note: Projected $142 million, actual $1.6 billion. When incentive costs balloon 11×, the welcome mat frays. EU municipalities courting hyperscalers should be running their own projections before signing.

Sources: Bloomberg (Alphabet), CNBC, AP News (Ohio), Fortune (Ohio)

AUKUS Launches Undersea Drone Programme to Guard the Cables Carrying the World’s Data

The US, UK, and Australia announced a joint programme to develop armed undersea drones with advanced sensors and weapons systems to protect submarine cables and energy pipelines. The project — the first major initiative under AUKUS Pillar Two — responds to what defence officials describe as an historically unprecedented surge in attacks on subsea infrastructure over the past 18 months. Deliveries are expected next year.

Note: Over 95% of intercontinental data crosses the ocean floor on cables thinner than a garden hose. When three defence ministries treat their protection as a military priority, that tells you what the intelligence assessments contain.

Sources: CNN

Frontier AI: The Race to Commodity

Opus 4.8 Leads the Intelligence Index — But Open Models Are Closing the Gap at 2% of the Price

The frontier keeps accelerating. Anthropic’s Opus 4.8, released May 28, leads the Artificial Analysis Intelligence Index at 61.4, edging out GPT-5.5 at 60.2. Early results on ARC-AGI-3 — a new benchmark for fluid adaptive intelligence where all systems scored below 1% of human efficiency as recently as March — suggest frontier models are beginning to crack tasks once considered far out of reach.

The open-weight pack is closing fast. Nvidia’s Nemotron 3 Ultra became the smartest open US model at roughly 550 billion parameters and an AAII score of 48, though China’s Kimi K2.6 still leads the open field at 54. MiniMax’s M3, launched June 1, claims to be the first open-weight model fusing frontier coding, a million-token context window, and native multimodality — at $0.12 per million input tokens against proprietary pricing above $5.

Note: A year ago, frontier AI meant choosing between two or three proprietary APIs. Today, open models cover most workloads at 2–3% of the price, and the intelligence gap is measured in single-digit index points. The cost floor is falling faster than most procurement cycles can track.

Sources: The Decoder (Nemotron), MiniMax (M3), VentureBeat (M3)

Institutional Productivity

Microsoft Is Building a Copilot Super App — One Surface for Chat, Code, Agents, and Automation

Microsoft is consolidating GitHub Copilot, Copilot Chat, Cowork, and a new agentic capability called “Autopilot” into a single application, with a launch planned by end of summer. The project, led by new Copilot head Jacob Andreou, responds to customer frustration with switching between scattered AI tools. A leaked screenshot also revealed a proactive agent called Scout — an anticipatory AI assistant that acts before being asked.

Note: When chat, coding, autonomous agents, and proactive suggestions collapse into one surface, the tool becomes the working environment. The hidden cost of that consolidation: a historian-turned-engineer recently warned that the mentor-to-junior handoff — the oral tradition that transferred institutional software knowledge — may not survive when the model handles both the work and the teaching.

Sources: Fortune, Fast Company

Capital Rotation

SoftBank Overtakes Toyota as Japan’s Most Valuable Company — An AI Bet Crowned Over the Combustion Engine

SoftBank surged 14% in Tokyo trading to reach ¥48 trillion in market capitalisation, overtaking Toyota’s ¥46 trillion and ending the automaker’s two-decade reign at the top. The last time SoftBank held the position was briefly during the 2000 internet bubble. The shift is almost entirely downstream of one bet: SoftBank’s ~13% stake in OpenAI, accumulated through commitments approaching $65 billion. Yesterday in this digest, SoftBank committed €75 billion to French AI data centres. Today, the market priced that conviction in.

Note: Japan’s most valuable company is now a holding company whose value rests on a single AI lab that hasn’t gone public yet. The market is pricing conviction, not diversification.

Sources: Financial Times, Nikkei Asia

More Than 220 Pre-ChatGPT Unicorns Are Now Stranded — Too Inflated for Venture, Too Unprofitable for Wall Street

PitchBook data published exclusively by CNBC reveals that more than 220 companies that once carried billion-dollar valuations are now “fallen unicorns” — cut off from venture funding by inflated price tags and undercut by AI competitors. Enterprise SaaS companies are hit hardest, with 75 of the 220 coming from that category alone. Startups that last raised in 2021 have lost 68% of their value on average; nearly half of the group hasn’t raised fresh funding in three years.

Note: The vendor ecosystem is consolidating in real time. For institutions relying on SaaS products from the 2019–2022 vintage, this is a supplier viability signal: your vendor may not raise another round.

Sources: CNBC


Today’s digest traces one arc from three angles: the capability to find ten thousand infrastructure vulnerabilities overnight, the capital to deploy $80 billion in a single equity raise, and the cost curve that puts frontier-grade AI at $0.12 per million tokens. The speed is the throughline. Institutions that planned on 18-month procurement cycles are operating in a market where the frontier moves in quarters, the toolchain consolidates in months, and an AI chatbot can hand over account credentials in minutes. The question is no longer whether to engage — it’s whether existing institutional timescales can keep pace with what’s already shipping.

Similar Posts