Welcome back to Level.UP, brought to you by UP.Labs.

This week, we look at what Anthropic's Claude Mythos means for the bridging layer between your factory and your cloud. And why Japan's physical AI bet should change how American operators think about their supply chain a decade out.

Plus: three tools worth your attention, AGIBOT's humanoids going live in a tablet factory, Antioch's bet to be the Cursor for physical AI, and R-Zero turning real-time occupancy into building operations.

Think someone else needs this? Forward it to a friend or colleague navigating the same terrain.

Anthropic recently announced Claude Mythos Preview, a powerful new model that it claims may be too dangerous to release to the public. 

The reasoning, confirmed by cybersecurity researchers: Mythos found thousands of zero-day vulnerabilities in internal testing, including a 27-year-old flaw in OpenBSD, an operating system specifically designed to be difficult to hack, and another in a line of code that had been tested five million times without detection.

Instead of a public launch, Anthropic gave roughly 40 companies access through Project Glasswing — including Microsoft, Apple, Google, Amazon, CrowdStrike, Nvidia, and others. The goal is to let defenders harden critical systems before a comparable model reaches attackers.

Nikesh Arora, the CEO of Palo Alto Networks, vividly described the attacker's side of the equation: imagine a horde of AI agents methodically cataloging every weakness in your technology infrastructure… constantly.

The question every operator should be asking about Claude Mythos is simple: how much of our industrial infrastructure is now protected by software that was never designed to be secure at scale?

The last decade of digital transformation connected the factory floor to the cloud, and the efficiency gains have been real: remote monitoring portals, edge gateways, data analysts pulling telemetry off PLCs, cloud dashboards that let a supplier in Stuttgart troubleshoot a line in Ohio. 

But it’s the server, not the PLC itself, that’s running the code Mythos was trained to hunt. The bridging layer — the IT/OT boundary — is where the next round of investment is needed.

There are two moves worth making now.

The first is assigning clear ownership of the bridging layer. Most security audits stop at IT or start at OT; the middleware in between is usually split across a plant manager, a CISO, and an outdated vendor contract. Put one person in charge, give them a budget, and let them rewrite the rules of engagement for every vendor that touches production data.

The second is self-auditing before someone else does it for you. Mythos will get easier to run the moment it (or something like it) reaches the open market. The operators who use that window to scan their own systems and remediate what surfaces will be in a different position than the ones waiting to find out what a bad actor found first.

Most operators will read this story as a warning. The sharper read is an invitation.

The bridging layer between IT and OT has been a source of drag for many companies for the past decade. Clarifying ownership could turn it into a source of speed. That's the trade Mythos is forcing, and the ones who take it will pull away.

On April 12, SoftBank, NEC, Honda, and Sony jointly incorporated Japan AI Foundation Model Development, a new company building a trillion-parameter AI model. The model isn't designed to write emails or generate images. It's designed to run factories, drive vehicles, and control robots.

The government is backing it. Japan's New Energy and Industrial Technology Development Organization has earmarked roughly ¥1 trillion, about $6.3 billion, in AI support over five years beginning in fiscal 2026. And the new company is considered a near-certain recipient.

The investor roster extends beyond tech: Nippon Steel, Kobe Steel, MUFG Bank, and Mizuho Bank have all taken stakes as well. Preferred Networks, the Tokyo-based deep learning specialist with existing ties to Toyota, is in for the R&D phase.

Japan's Ministry of Economy, Trade, and Industry aims to capture 30% of the global physical AI market by 2040, building on an existing position where Japanese manufacturers account for roughly 70% of the global industrial robotics market. Their target for real-world deployment is 2030.

The most underpriced asset in this consortium is the training data.

Japanese industrial robots are installed in thousands of factories worldwide, spanning automotive, electronics, semiconductors, food processing, and pharmaceuticals. Every one of those deployments generates operational data about how physical systems actually behave under production load. 

That data isn't available to Nvidia, Tesla, or OpenAI. A foundation model trained on it doesn't just rival what Nvidia and Tesla are building; it could block American entrants from physical AI markets they assumed were theirs to win.

The government posture signals that this is not a venture bet. Tokyo is standing up a sovereign wealth fund and an AI consortium in the same quarter, with the country's largest steelmakers and three largest banks taking equity alongside the tech companies.

Japan is treating physical AI the way it treated consumer electronics in the 1980s and automotive manufacturing in the 1990s. And we’ve seen that playbook work before.

The SoftBank question is the one American operators should be paying attention to. Masayoshi Son has been the single largest source of capital for US AI dominance for nearly a decade. He led OpenAI's $40 billion round last year.

Now he's anchoring a Japanese alternative explicitly designed to reduce dependency on the same American stack he's been funding. If even a fraction of that capital flows to Tokyo, the funding base for US physical AI would look materially different.

The harder question is whether the US has any meaningful time advantage left.

China has a manufacturing base. Japan has the installed robotics base, the capital, and now the coordinated plan. Both countries are positioned to extend their manufacturing dominance into foreign factories by selling the operating systems that run them.

The US has the models and the chips. Whether that's enough to win a market where the moat is operational data from decades of deployment is an open question, and it's being answered in real time.

Ready to work smarter? Here are the tools we're using to actually get more done:

R-Zero Just Connected Room-Level Occupancy To Real-Time Building Operations. The San Francisco-based physical AI platform launched Room Intelligence within its Connect product this week, turning live occupancy data into automatic decisions about ventilation, energy use, and space planning. Customers include pediatric healthcare systems and a Fortune 100 company with 79,000 employees across seven offices. The signal: the physical AI opportunity in built environments isn't just in new sensors — it's structuring the data [that's already there] around how facilities teams actually run buildings. → Read more

Antioch Raised $8.5M To Be The "Cursor For Physical AI." The New York startup, valued at $60 million post-money, is building simulation tools that let robot developers spin up digital twins of their hardware and test against realistic sensor data before touching physical prototypes. The pitch: every autonomous system will be built primarily in software within two to three years, and the sim-to-real gap is the bottleneck. → Read more

AGIBOT Is Deploying Humanoid Robots Into A Live Tablet Factory. The Shanghai robotics company integrated its G2 humanoids into Longcheer Technology's mass-production line for consumer electronics — producing 310 units per hour, a 99.9%+ success rate, and full integration in 36 hours. AGIBOT plans to scale to 100 robots by Q3 and move into automotive, semiconductors, and energy next. It’s the first large-scale industrial deployment of embodied AI inside core manufacturing workflows. The signal: China isn't waiting for the humanoid debate to be settled. Embodied AI is exiting the demo phase. → Read more