Meet Your AI Workforce

This week, Anthropic built an AI so powerful they refused to release it. Instead, they handed it to Apple, Microsoft, and Google to hunt for security flaws. Meanwhile, an open-source project is letting people run entire companies staffed by nothing but AI agents. The future of work just got very real.

The Big Story

Anthropic locked down its most powerful AI model, then put it to work hunting zero-day vulnerabilities

Anthropic quietly unveiled Claude Mythos, a new frontier model so capable at finding software vulnerabilities that the company decided not to release it publicly. In internal testing, Mythos discovered thousands of previously unknown zero-day flaws across every major operating system and web browser, including a 27-year-old bug in OpenBSD that had gone undetected for nearly three decades.

The striking part: Anthropic did not train Mythos to do this. These cybersecurity capabilities emerged on their own as a side effect of improvements in coding and reasoning. That should make you pause. When an AI gets good enough at understanding code, it starts seeing weaknesses that humans have missed for decades.

Rather than shelving the model, Anthropic launched Project Glasswing, a $100 million initiative giving roughly 40 organizations limited access to Mythos for defensive security research. The partner list reads like a who's who of tech: AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and NVIDIA.

My take: This is the most responsible AI deployment I have seen from any frontier lab. Instead of racing to ship the model and deal with consequences later, Anthropic chose to restrict access and channel its capabilities toward defense. If you use any major browser or operating system (you do), the software you rely on every day is about to get meaningfully more secure because of this decision. The tradeoff is that the rest of us do not get to play with Mythos. For now, that seems like the right call.

Tool of the Week: Paperclip

What it does: Paperclip is an open-source platform that lets you organize AI agents into a company structure, complete with org charts, roles, budgets, and task delegation, so they can run business operations autonomously.

Why it matters this week: Paperclip launched on March 2, 2026 and has already crossed 51,000 GitHub stars, making it one of the fastest-growing open-source AI projects of the year. On April 3, a developer named @dotta stood up an entire content company called Effloow staffed by 14 AI agents using Paperclip, with a CEO agent delegating work to manager agents who assign tasks to worker agents.

How to use it:

1. Install with one command: npx paperclipai onboard --yes
2. Define your "company" by setting a mission and creating agent roles (CEO, content writer, researcher, etc.)
3. Connect your preferred AI agents (works with Claude, OpenClaw, Cursor, Codex, or any HTTP-based agent)
4. Set heartbeat schedules (every 5, 30, or 60 minutes) so agents wake up, check for work, act, then go dormant
5. Monitor everything from the dashboard: live status, cost tracking, decision logs, and task progress

The real power is in the governance layer. Every agent gets a monthly budget with hard limits, every decision is traced and logged, and you can set approval gates for high-stakes actions. It is not a chatbot on autopilot. It is a managed workforce with accountability built in.

Pricing: Free and open-source (MIT license). Self-hosted. Your only cost is the API usage of whatever AI models you connect.
Link: paperclip.ing

Quick Picks

• Microsoft MAI Models: Microsoft launched three in-house AI models: MAI-Transcribe-1 (beats Whisper in all 25 languages), MAI-Voice-1 (60 seconds of natural audio in one second), and MAI-Image-2. The message is clear: Microsoft is building its own AI stack and reducing its dependence on OpenAI.
• Gen Z AI Skepticism Climbs: Gallup survey finds 51% of Gen Z use AI weekly, but excitement dropped 14 points to just 22%, while anger rose 9 points to 31%. Usage is holding steady; trust is not.
• Apple Building AI Agents with Guardrails: Apple is designing AI agents that require human approval before taking sensitive actions like making payments. A reminder that "autonomous" does not have to mean "unsupervised."