Anthropic built an AI that hacks. Then gave it to defenders first.

iPrompt
Friday Edition

…all in a FREE Weekly newsletter.

AI & robotics investing — explained so you can actually act on it.  |  10 April 2026  |  Friday Edition 

THE HOOK

Somewhere in San Francisco this week, a researcher at Anthropic received an email from an AI model. He was eating a sandwich in a park. The model had broken out of its sandbox, built a multi-step exploit, and decided to tell him about it. That model is called Mythos. It can break into every major operating system on earth. Anthropic isn’t selling it. It’s handing it to 40 companies and saying: patch everything.

The cybersecurity sector just got repriced.

WEEKLY SCOREBOARD

VIX = the ‘fear gauge.’ Measures expected S&P 500 swings over 30 days. Below 20 = calm. It’s been above 20 for six straight weeks on Iran. Ceasefire pulled it back, but we’re not out yet.

Bottom line: Ceasefire gave markets permission to rally. But the real repricing happened in cybersecurity — CRWD and PANW moved harder than the index, on a completely different catalyst. That’s the tell.

TOP HEADLINES

1. Anthropic launches Project Glasswing, restricts Mythos to 40 organisations

Anthropic’s new frontier model found thousands of zero-days — some decades old — across every major OS and browser. Rather than release it publicly, Anthropic handed Mythos to AWS, Apple, Microsoft, CrowdStrike, and others, backed by $100M in credits and $4M for open-source security. First time a frontier lab has withheld a model over offensive capability since OpenAI delayed GPT-2 in 2019. The result is a two-tier AI access structure: companies inside the defensive perimeter, and everyone outside it.

2. Meta kills its own open-source narrative

The company that spent three years evangelising open AI just shipped a proprietary model. Muse Spark, built by Alexandr Wang’s Superintelligence Labs, is closed-source, API-restricted, and powers Meta AI across all apps. Benchmarks put it competitive with Opus 4.6 and Gemini 3.1 Pro. Meta’s AI capex for 2026: $115–135 billion. That’s nearly double last year. The moat for closed-model companies just got wider, not narrower.

3. OpenAI, Anthropic, Google share threat intelligence on Chinese distillation

Three companies that compete for the same enterprise contracts are now sharing data through the Frontier Model Forum. The targets: DeepSeek, Moonshot AI, MiniMax. Anthropic documented 16 million unauthorised exchanges from ~24,000 fake accounts. When rivals cooperate like this, the threat isn’t commercial. It’s something else entirely.

4. Q1 venture: $300B, and 65% went to four companies

Crunchbase data: $300 billion into 6,000 startups globally in Q1, up 150%+ from Q4. OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B) closed four of the five largest rounds ever recorded. AI accounted for 80% of total global venture funding. This isn’t a broad startup boom. It’s capital concentration at a scale that makes the dot-com era look diversified.

5. Japan’s workforce is shrinking. Its robot output isn’t.

METI announced a national strategy targeting 30% of the global physical AI market by 2040. Japanese manufacturers already make roughly 70% of the world’s industrial robots. But here’s the bit that makes this different from every other government AI announcement: Japan isn’t choosing to automate. Demographics are forcing it. The workforce contracts every quarter. For investors, that’s the most compelling non-US AI thesis available right now — necessity-driven, not narrative-driven.

OUR INVESTING ANGLE

Everyone’s watching the AI model race. The smarter bet is watching who profits from the fallout.

The thesis: AI just split the cybersecurity market in two. Companies inside the Glasswing perimeter get AI-grade defence, early vulnerability data, and the institutional bid that comes with being named critical infrastructure. Companies outside it become softer targets by comparison. The gap between those two groups is where the next 12 months of security alpha lives.

I’ve been watching cybersecurity since the Mythos leak in March, and this week was the cleanest thesis reversal I’ve seen in a while. Follow the logic:

• Mythos finds zero-days in every major OS and browser (headline 1). The attack surface got bigger this week. Permanently.

• The Frontier Model Forum goes operational (headline 3). When three companies that fight over the same enterprise contracts share data willingly, you know something crossed a line.

• 80% of Q1 venture capital flowed into AI (headline 4). Every dollar into capability is a dollar that eventually needs defending.

The market initially read Mythos as a threat to security vendors. That was wrong. Glasswing flipped the narrative: CRWD and PANW aren’t being replaced. They’re being wired into a new AI security architecture as the defensive layer. JPMorgan called them ‘essential layers in the defensive stack, not competitive targets.’

But here’s the part nobody’s discussing. The losers. Think Fortinet, SentinelOne, Carbon Black — solid companies, decent products, real client bases in logistics and healthcare. None of them were on the Glasswing list. They don’t have AI-native detection. They’re still selling signature-based tools built for a pre-Mythos threat landscape. Their customers are now, objectively, less protected than a CRWD or PANW client. And the CISO at those customer companies is about to find that out at their next board meeting.

So yeah. That’s the trade.

The thesis isn’t ‘cybersecurity benefits from AI.’ Everyone sees that now. The thesis is that the benefit lands unevenly — and the names on the Glasswing list are the ones capturing the premium.

Read the full analysis: [The AI Security Paradox — deep dive on iprompt.com]

THREE IDEAS TO RESEARCH THIS WEEKEND

Not recommendations — starting points for your own research.

IDEA 1: The Defensive Perimeter Trade

Glasswing named CRWD and PANW as founding partners. JPMorgan reiterated overweight on both. Both sit at the intersection of AI and security spend — PANW at 25% discount to JPMorgan’s $213 target, CRWD coming off its best day in six months. Both still down on the year, which means entry sits below pre-Mythos levels.

Risk: Valuations still stretched — PANW at 48x, CRWD higher. A broad pullback erases the Glasswing premium in a session.

Tripwire: If CRWD earnings (1 June) show net new ARR below 50% YoY, the growth premium is unjustified. If Glasswing expands beyond 100 partners within 90 days, the exclusive-perimeter thesis dilutes.

How to research: CRWD, PANW. Broader: HACK ETF, CIBR ETF.

IDEA 2: Japan Physical AI — Necessity Beats Narrative

What if the most interesting AI trade right now isn’t in the US? Japan’s METI wants 30% of the global physical AI market by 2040. Japanese manufacturers already make 70% of the world’s industrial robots. The workforce is shrinking every quarter — not choosing to automate, being forced to. FANUC and Keyence trade at meaningful discounts to US AI valuations. I keep coming back to this one.

Risk: Yen weakness erodes returns for non-Japanese investors. Government targets ≠ commercial outcomes.

Tripwire: If BYD or Huawei enters industrial robotics with sub-cost pricing by Q3, the Japanese valuation gap is structural, not a mispricing. Watch Keyence’s next quarterly for factory automation orders.

How to research: BOTZ ETF (43.8% industrials, heavy Japan), FANUC (6954.T), Keyence (6861.T).

IDEA 3: Meta’s Proprietary Pivot

Look, I’m not entirely sold on this one. But the setup is hard to ignore. Meta launched Muse Spark as a closed model this week — paid API preview, Shopping mode, $115–135B in capex. That’s the Anthropic/OpenAI playbook, run by the company with 3 billion daily users and an advertising machine that prints money. At 22x forward P/E, you’re getting the ad business almost free if the AI monetisation works.

Risk: Muse Spark trails on coding and agentic tasks. Meta’s privacy reputation is a real headwind for enterprise.

Tripwire: If Meta’s Q2 earnings call (late July) doesn’t break out AI API revenue as a separate line item, the monetisation thesis stays theoretical. If Muse Spark user retention in the Meta AI app falls below 30-day figures for ChatGPT or Claude, the product isn’t sticky enough to justify the capex.

How to research: META. Compare MSFT, GOOG for platform AI.

AI INVESTMENT FRAMEWORK

6 layers · Updated weekly · Not financial advice

■ Infrastructure

AI capex cycle is early innings. Every major cloud provider is increasing spend.

NVDA AI GPU dominance. Earnings 20 May.  AVGO Custom chips + networking. $100B AI rev target FY27.  MU HBM4 demand validated.

■ Platforms

Platform layer captures value regardless of which model wins.

MSFT Azure AI + Copilot. Most diversified AI play.  GOOG Gemini 3.1. Vertex + Anthropic partnership.  AMZN AWS Bedrock. $50B OpenAI anchor.

■ Applications

Hype-to-revenue transition. Winners emerge after Q1 earnings cycle.

PLTR Down 18% YTD. Valuation compression despite 70% rev growth.  CRM Agentforce rollout.  NOW AI workflow automation.

■ Physical AI

Where software meets the real world. Japan leads on demographics.

BOTZ 43.8% industrials, heavy Japan. Lowest vol in framework.  ISRG Surgical robotics leader.  FANUC Japanese industrial robotics. METI tailwind.

■ Cybersecurity  NEW

AI creates threats only AI can defend. Glasswing partners are the defensive perimeter.

CRWD Glasswing founding partner. +6.2% this week. Earnings 1 June.  PANW JPMorgan top pick. $213 target.  ZS Zero-trust leader.

■ Global

Non-US AI exposure hedges concentration risk. Geopolitics is the wildcard.

BABA Qwen models competitive. Deep discount.  9984.T SoftBank. ARM + OpenAI stake.  SAP European enterprise AI. Joule copilot.

Equal-weighted layer indices, rebased to 100 on 2 January 2026. Cybersecurity (red) cratered in March on the Mythos leak, then snapped back on Glasswing. That V-shape is this week’s thesis in one line.

WHAT WE’RE WATCHING

Changes this week: Added Cybersecurity as a standalone layer at HIGH conviction. Glasswing isn’t a one-off — it names these companies as critical infrastructure in the AI security stack. Previously a sub-theme under Applications. This week justified the promotion.

Disclaimer: This newsletter is for informational and educational purposes only and does not constitute financial advice. iPrompt Signals is not a registered investment advisor. Always conduct your own research and consult a qualified financial professional before making investment decisions.

YOUR MOVE

Three things from this week:

1. AI weaponised cybersecurity — both offence and defence. The companies in Glasswing aren’t in a marketing exercise. They’re building the defensive layer for a threat landscape that arrives in months, not years.

2. The Frontier Model Forum stopped being a talking shop. It’s an operational intelligence alliance now. When OpenAI, Anthropic, and Google share data willingly, something crossed a line.

3. Japan’s physical AI push is the thesis nobody’s pricing. Shrinking workforce, government backing, 70% of global robot manufacturing. Setup, not story.

Now research one.

Which catalyst matters most for your portfolio this quarter? Hit reply — the best responses make it into next week’s P.S.

Stay curious — and stay qualified.

— R. Lauritsen

Know someone building an AI position? Forward this — they’ll thank you by Friday.

P.S. Two weeks ago we flagged cybersecurity as oversold after the Mythos leak. Glasswing confirmed it. If you acted on that — nice one. If not, entry’s still below February levels. Which, fair enough, is annoying to hear.

QUICK GLOSSARY

Adversarial distillation: Copying another AI’s capabilities by querying it millions of times and training on the answers. Fine when you do it to your own models. Not fine when a competitor does it to yours.

Capex: Capital expenditure — money spent on infrastructure, chips, servers. When Meta says $115–135B, that’s what staying in the AI race costs.

Frontier model: The most capable model a given lab has built. Mythos is Anthropic’s. Muse Spark is Meta’s.

Net new ARR: Annual Recurring Revenue added in a quarter, minus churn. The growth number that matters for SaaS businesses like CrowdStrike.

VIX: CBOE Volatility Index. Expected S&P 500 swings over 30 days. Above 25 = nervous. Below 15 = too comfortable. Currently ~21.

Zero-day: A software flaw nobody knows about yet. No advance warning, no patch. The dangerous kind.