Anthropic Built It. Then Hid It.
Issue #9 A model that can break into any system. And you will never get access to it.
Hey Folks, It was a strange week.
because…
Anthropic built something that can break into real systems. Then decided nobody gets to use it.
That one call sparked more conversation than any model launch this year.
A few other things quietly happened this week, too.
OpenAI is chasing ad money.
Agents hit a milestone that surprised many people.
Anthropic is suddenly worth $380 billion.
Lot to cover. Let’s go.
The AI Story That Mattered This Week
Anthropic built a hacking AI model and locked it away
Anthropic quietly completed a new model called Claude Mythos.
The unusual part is they are not releasing it.
Not because it was unfinished. Because it was too capable at something they did not expect.
Mythos can identify and chain real software exploits across major operating systems, browsers, and the Linux kernel. It discovered a 27-year-old bug in OpenBSD. It operates with minimal human guidance. In controlled tests, it succeeded more than 80% of the time.
That number changes the conversation.
Previous frontier models could find two working exploits in similar tests. Mythos identified vulnerabilities across thousands of production systems.
Anthropic’s response was to restrict it entirely. Access goes only to a small group called Project Glasswing. Partners include Amazon, Microsoft, Apple, Google, and Nvidia. The goal is defensive research only. Anthropic committed $100 million in usage credits and coordinated directly with governments to patch vulnerabilities before they spread.
One uncomfortable detail stood out in expert discussions. If Anthropic built this, other labs probably already have something similar.
Security research may never look the same again.
THREE SIGNALS FROM THIS WEEK
AI agents jumped from 12% to 66% task success on real computer tasks in one year
Anthropic overtook OpenAI in enterprise API market share, 40% versus 27%
OpenAI is pivoting toward advertising as its next major revenue stream.
HEADLINES
OpenAI wants to become an ad business. OpenAI is projecting $2.5 billion in ad revenue this year. By 2030, they want $100 billion annually from ads alone. An early pilot hit $100 million annualized in two months. They are using what you type in chat to show relevant ads. For a company that started as a research lab, this is a big identity shift.
Agents just crossed a line; nobody expected this fast. One year ago AI agents succeeded at 12% of real computer tasks. Today that number is 66%. Web traffic from autonomous agents is up 7,851% year over year. Gartner says 42% of companies will deploy agents in the next 12 months. The same report says 40% of those projects will fail by the end of 2027.
Most security leaders have no plan if an agent goes wrong. 86% of CISOs have zero access policies for AI agents running inside their companies. Only 5% think they could stop a compromised agent. This is happening right now while agent deployments scale into thousands across enterprise environments. Nobody is ready for this.
Anthropic ignores $800B takeover-style funding pressure as growth explodes. Anthropic is being offered VC deals valuing it near $800B, but is not rushing in. The company previously raised at a $380B valuation after a $30B round. Since then, revenue reportedly hit $30B annualized, driven by enterprise demand and coding tools like Claude Code. Over 1,000 customers now spend $1M+ yearly. Claude Code alone runs at ~$2.5B revenue scale. OpenAI’s enterprise share has slipped as competition intensifies.
PRODUCT UPDATES
Claude can now use your computer Anthropic shipped Computer Use for Claude. It can browse, open files, click through interfaces, and finish multi-step tasks on its own. Not just answering questions anymore. It is actually doing things. Developers and enterprise teams are the first to get access.
HubSpot stopped charging monthly fees HubSpot switched to outcome-based pricing. $0.50 per resolved conversation. $1 per qualified lead. You pay nothing if the AI does nothing useful. Every SaaS company still charging flat fees for AI features should be paying attention to this.
Anthropic’s Advisor Tool lets cheap models ask smart ones for help Haiku can now pause mid-task and consult Opus 4.6 when it hits something hard. In testing, this doubled Haiku’s SWE-bench score from 19.7% to 41.2% at 85% lower cost. Good quality at a fraction of the price. That is a meaningful shift for developers watching their API bills.
AWS built a control tower for your agents. Amazon launched Agent Registry this week. One place to see and manage all your AI agents. Most companies running agents right now have no central visibility into what those agents are actually doing. This is the first serious attempt to fix that.
Cloudflare made agent deployment faster. Cloudflare launched Mesh; AI agents can now connect to private company networks in minutes instead of days. One of the biggest friction points in enterprise agent work just got a lot smaller.
TOOLS TO USE THIS WEEK
Google Fabula helps you structure and refine your writing instead of just generating more of it. Good if your ideas are solid but your organization is a mess.
Amazon Bio Discovery Lets scientists run complex biological workflows and generate molecules without writing a single line of code. Built for early-stage drug discovery teams.
PaperOrchestra Turns raw research notes into structured papers using multiple AI agents working together. Useful if you are drowning in information and need something readable fast.
Cognee gives AI agents real memory across sessions using graph, vector, and relational storage. If you are building agents that need to remember things, start here.
NEW DEVELOPMENTS
Five companies own most of the world’s AI compute: Google, Microsoft, Meta, Amazon, and Oracle control over two-thirds of global AI compute. Every AI lab building frontier models depends on one of these five companies. That kind of concentration is a problem the industry has not seriously talked about yet.
AI is writing most of the code now; 80% of developer teams use AI coding tools. Code acceptance rates went from 20% to 60% in one year. AI leads development workflows at most serious engineering teams. The real question now is: How much of the final product did a human actually write?
A research paper did the math on AI layoffs: UPenn and Boston University researchers proved with game theory that AI-driven layoffs follow a Prisoner’s Dilemma. Each company wins short-term by cutting jobs. If everyone does it, consumer spending collapses. Their fix is an automation tax on companies replacing workers with AI. 100,000 tech jobs disappeared in 2025 alone.
If this helped you understand the week in AI more clearly, share it with someone who follows the space closely.
Before you go, please give me your feedback below. It helps me shape every issue.
See you next Thursday.