Mythos 1 , neocloud boom , MCP goes stateless
Mythos 1, neocloud boom, MCP goes stateless
Three moves that, on their own, look like industry news — and together, look like the architecture of the next 18 months. Anthropic's Claude Mythos is graduating from a closed preview to a wider defensive network. The "neocloud" GPU-rental market is hitting the capex wall and the unit-economics wall at the same time. And the Model Context Protocol — the standard every agent builder is now betting on — is going stateless for hyperscale.
What You Need to Know: The European Central Bank has called its supervised banks into a closed meeting to coordinate around Claude Mythos, the Anthropic model that has reportedly produced working zero-day exploits on first attempt more than 83% of the time. Synergy Research says the neocloud market is on track to approach $400B by 2031, with a 58% CAGR — and that is the floor, not the ceiling. And the MCP maintainers, including David Soria Parra at Anthropic, are openly mapping a stateless redesign for hyperscale, plus a long-running-tasks model that fixes the protocol's biggest production pain point.
Why It Matters
- A frontier model that writes working exploits faster than defenders can patch is not a tool. It is a new layer of national infrastructure, and the ECB is treating it as one.
- If you build on AWS, Azure or GCP today, your margin is being squeezed by hyperscaler capex. If you build on a neocloud, your margin is being squeezed by token prices. The choice is now architectural, not tactical.
- MCP going stateless is the boring news that fixes the largest production headache in agent infrastructure. If you have ever lost session state mid-task, this is your upgrade path.
- "Long-running tasks" is the formalization of what every agent builder has been duct-taping together with cron jobs and queue workers. The standardization is overdue.
- Mythos is still gated. The capability is not. Defenders need to plan on the assumption that the model — or something like it — is already in the hands of the adversary.
What Actually Happened
Claude Mythos Steps Out of Preview, Into European Bank Defenses
The European Central Bank is convening European banks on a Tuesday in late May to deal with the cybersecurity implications of Claude Mythos Preview, the gated Anthropic model that, according to reporting in the Financial Times, has identified thousands of zero-day vulnerabilities across major operating systems and browsers. ECB executive board member Frank Elderson told the FT that the window between a vulnerability being patched and being exploited has "collapsed" — AI models can now reverse-engineer software fixes within minutes of their release.
Only about 40 to 50 organisations currently have access to Mythos through Anthropic's controlled-distribution programme, called Project Glasswing. The list includes Amazon, Microsoft, Google, Nvidia, CrowdStrike, Palo Alto Networks and JPMorgan Chase. No European bank is on it. In controlled testing, the model produced working exploits on its first attempt more than 83% of the time, often outperforming human security researchers. Anthropic has warned that adversaries could replicate the capability within 6–12 months. (thenextweb.com)
The Neocloud Boom Hits Its Stride
The "neocloud" tier of GPU-rental providers — CoreWeave, Lambda, Crusoe, Nebius, the smaller neoclouds you have never heard of, plus the regional tier-2s — has gone from a 2023 niche to the second-largest infrastructure bet of the cycle. Synergy Research's April forecast puts the market on track to approach $400 billion by 2031, a sustained 58% CAGR. One substack estimate puts CoreWeave's run-rate revenue at $3.4 billion on a fleet scaling from 23,000 to 140,000 GPUs in 2026, with the fleet still running at only 16% of nameplate utilization. That utilization figure is the entire story: the capex is already spent. The pricing power is not. (srgresearch.com, sergeycyw.substack.com)
MCP Goes Stateless, Adds Long-Running Tasks
The Model Context Protocol — Anthropic's open standard for connecting agents to data and tools, and the de facto transport layer for every serious agent framework shipping in 2026 — is getting a structural overhaul. David Soria Parra, one of the protocol's maintainers at Anthropic, walked the 2026 roadmap publicly, and the two biggest items are a stateless redesign for hyperscale deployments and a formal long-running-tasks model. The stateless piece fixes the most-cited production bug: a developer attempting to build a stateless MCP server across multiple pods, with state stored in Redis, has been fighting a long-standing GitHub issue about coordination since last summer. The long-running-tasks model formalizes what agent builders have been doing with cron, queue workers and bespoke checkpointing — except now the protocol owns the contract. (youtube.com — David Soria Parra, AI Engineer Summit, thenewstack.io)
Mythos and the Defensive Window
Palo Alto Networks has been tracking this in the wild. Their data, referenced in the ECB's reporting, shows advanced AI models are discovering vulnerabilities at seven times the usual rate. The firm has warned the industry has only 3–5 months of defensive buffer remaining. Anthropic has briefed the Financial Stability Board on what Mythos has been finding, at the request of Bank of England governor Andrew Bailey. The Federal Reserve and the US Treasury have separately convened bank CEOs. This is not a vendor pitch. It is a regulatory event. (thenextweb.com)
The Take
Three different industries, one shared pattern. Mythos is forcing banks to defend against a model they cannot use. The neoclouds are growing into a tier that will rival AWS by 2030 on raw GPU capacity, but with no margin floor. And MCP is fixing the production gaps that have been the dirty secret of agent infrastructure since 2024.
If you are building on top of any of these, here is the play. Treat Mythos — or its inevitable open replication — as a baseline threat model. It will be in the hands of attackers within a year. Build the neocloud decision the way you would build the cloud decision in 2015: workload-specific, not provider-wide. And if you are running MCP in production, start designing for the stateless migration now, because the upgrade is going to be the kind of breaking change you want finished before your traffic doubles.
The boring story in this digest is the MCP one, and it is the one that matters most for builders. The Mythos story is the loud one, and it is the one your security team needs to be planning around this quarter.
Quick Summary
Anthropic's Claude Mythos is graduating from gated preview into the European bank defensive playbook, with the ECB convening lenders around a model that produces zero-day exploits on first try 83% of the time. The neocloud market is on a path to $400B by 2031. And MCP is going stateless and adding a long-running-tasks model — the two changes that will unblock agent infrastructure in production.