Claude Mythos found a 27-year-old bug in OpenBSD that had survived decades of human review. It found a 16-year-old bug in FFmpeg that automated testing tools hit 5 million times without catching. It chained together Linux kernel vulnerabilities to get root access from an ordinary user account.

All of it was autonomous. No human involvement after the initial prompt.

Instead of releasing Mythos to the public, Anthropic assembled Apple, Microsoft, Google, Amazon, CrowdStrike, and 40 other organizations into something called Project Glasswing: a $100 million coalition to patch the world's infrastructure before models like this proliferate.

This is the first time Anthropic has published a system card for a model without releasing it. The cyber capabilities are too dangerous for public access.

Meanwhile, the pattern from last week continues: the AI tools winning are the ones that connect to your actual work. Mythos connects to codebases. Claude Code connects to your terminal. Cowork connects to your desktop files. MCP connects everything else.

This week: how to actually set up Claude Cowork from scratch, and why Mythos matters even though you can't use it.

The Old Way:

You download Claude Desktop. You click around. You're not sure what Cowork actually does or why it's different from regular Claude. You try a task, get mediocre results, and go back to chat mode.

Meanwhile, people on Twitter are showing Claude organizing their entire file systems, generating weekly reports on autopilot, and running scheduled tasks while they sleep. You wonder what you're missing.

The Replacement:

A step-by-step system that sets up Cowork properly from the start. Not just installation, but the folder structure, identity files, global instructions, and scheduled tasks that make Cowork actually useful.

Cowork isn't a chatbot with file access. It's a desktop agent powered by Claude Opus 4.6 with a 1 million token context window. It reads local files, executes multi-step workflows, and delivers finished outputs directly to your file system. But it only works well if you set it up right.

Windows support launched February 10, 2026. Feature parity with Mac. The addressable audience doubled, but the setup guides didn't get better. This system fills that gap.

In regular Claude chat, you upload files manually and copy-paste outputs. Context resets every conversation. You're the middleman between Claude and your work.

In Cowork, Claude has direct access to folders you specify. It reads what's there. It creates new files. It runs multi-step tasks without you babysitting. And with scheduled tasks, it can do work while you're not even at your computer.

The difference in practice: regular Claude helps you work. Cowork does work.

But the setup matters. Most people skip it, then wonder why their results are generic.

This system takes you from fresh install to functional automations in about 60 minutes.

Prompt 1: Workspace Architect: Creates your folder structure and organizes your Cowork environment

Prompt 2: Identity Builder : Generates your about-me.md and anti-ai-style.md files so Claude knows how to sound like you

Prompt 3: Global Instructions Writer: Creates the settings Claude reads before every task

Prompt 4: First Project Creator: Sets up a recurring work project with scoped memory

Prompt 5: Scheduled Task Builder: Automates a weekly deliverable that runs while you sleep

First month is $9. Cancel anytime.

Cursor 3: Cursor rebuilt their entire IDE around AI agents. The new "Agents Window" lets you run multiple agents in parallel across repos and environments. Assign tasks to agents, monitor their progress, merge their outputs. Direct competitor to Claude Code but with a visual interface.

Algrow YouTube MCP: Claude can now search YouTube with real data. Find competitors, pull video performance metrics, grab transcripts, analyze thumbnails. 19 tools in one connector. Solves the problem that Claude's training data knows nothing about what's actually working on YouTube right now.

Cohere Transcribe: Open source speech-to-text that runs on consumer GPUs. 2B parameters, 14 languages, 5.42% word error rate (best in class). Apache 2.0 license. Good for anyone building transcription into products without paying per-minute API fees.

Yahoo Scout: Yahoo's AI answer engine, powered by Claude. Not a chatbot. Direct answers with citations, integrated into Yahoo Mail, Finance, Sports, and News. Available to 250 million U.S. users. Interesting because it shows Claude powering mainstream consumer search.

Axon Vision: AI for live video analysis, launched at Axon Week 2026. Detects critical activity in CCTV feeds. Enterprise/law enforcement focused, but signals where computer vision is heading: real-time analysis, not just post-processing.

🔐 Anthropic announces Claude Mythos and Project Glasswing. Most powerful model ever built. Too dangerous to release. Found thousands of zero-day vulnerabilities in every major OS and browser. Apple, Microsoft, Google, Amazon, and 40 other orgs get access for defensive security work. $100M in credits for the coalition. (Anthropic, Fortune, CNBC)

⚙️ Cursor 3 launches with agent-first interface. Complete rebuild around parallel AI agents. Cloud and local execution, Design Mode for visual UI editing, Composer 2 model with high usage limits. Direct competition with Claude Code. (Cursor, Changelog)

🎯 Google drops Veo 3.1 pricing. Veo 3.1 Lite launched at less than 50% the cost of Fast. Veo 3.1 Fast gets a price cut on April 7. Video AI getting cheaper as compute efficiency improves. (Google Blog, The Decoder)

📊 Tech layoffs hit 52K in Q1 2026. Atlassian cut 1,600 employees (10% of workforce) to redirect to AI. Block, Amazon, Meta, and others followed. The pattern: cut headcount, invest in AI tools. Workforce restructuring around automation is accelerating. (Bloomberg, TechCrunch)

What happened:

On April 7, 2026, Anthropic announced Claude Mythos Preview and Project Glasswing in the same breath. One is the most capable model they've ever built. The other is the reason they're not releasing it.

Mythos is a general-purpose frontier model, not specifically trained for security. But its coding and reasoning capabilities turned out to be so strong that it started finding vulnerabilities human security researchers had missed for decades.

The highlights:

Instead of releasing it, Anthropic assembled Apple, Microsoft, Google, Amazon, and 40 other organizations into Project Glasswing: a $100 million coalition to patch the world's infrastructure before models like this proliferate.

This is the first time Anthropic has published a system card for a model it won't release.

🔒 The full breakdown is for paid members: how Mythos chains vulnerabilities together, why Nicholas Carlini says he found more bugs in two weeks than his entire career, what this means for your work, and implementation steps.

@AnthropicAI — "Today we're announcing Project Glasswing and Claude Mythos Preview. We're committing $100M in usage credits and bringing together 12 major organizations to secure the world's most critical software." Read the full announcement

@simonw — "The surrounding code is from 27 years ago. I tracked the change down in the GitHub mirror of the OpenBSD CVS repo (apparently they still use CVS!) and found it using git blame." Simon verifying the vulnerability timeline. Read his analysis

@cursor_ai — "We're introducing Cursor 3. It's simpler, more powerful, and built for a world where all code is written by agents." The agent-first IDE launches. Read the announcement

@ilovetechlit — "Two-man AI startup hits $401 million in first year. Matthew Gallagher built Medvi with ChatGPT, Claude, Midjourney, and Runway. One employee (his brother). No outside funding. NYT verified the numbers." The most 2026 headline possible. Read the story

🔒 Implementation checklist is for pro members. Get a day-by-day plan to implement AI into your workflow.

Personal Brand Starter Kit

Prompt Library Updates:

Resources This Month:

That's the state of AI this week. Anthropic built a model too dangerous to release. The security industry just changed. And the tools that actually help you work (Cowork, Claude Code, MCP) keep winning while the flashy demos die.

Set up Cowork properly. The 60 minutes you spend now saves hours every week.

See you next week.

Weekly:

Ongoing:

🎉 First month is $9. Cancel anytime.