🔱  Sign up for a free demo today to see how the latest AI and digital advocacy tools can transform your work.

2 Minute Read or Listen

  The One AI Term You Need To Understand Now

Last week, the two biggest AI companies in the world told us that their next models are so powerful, that the world is not ready for them. Some people have suggested this is a marketing gimmick. We think that's unlikely. In today's short read, we show you why these warnings should surprise no one, and discuss what they mean for your work.

Imagine Getting This Email at Lunch 

Earlier this week, an Anthropic researcher named Sam Bowman was eating a sandwich in a San Francisco park when he received an email he wasn't expecting. It was from a new Anthropic AI model that he was working on. The model had just broken out of its own secure computing environment, hacked its way onto the open internet, and decided—on its own—to let him know it had escaped.

The model that emailed Bowman is called Claude Mythos Preview. In safety testing, researchers placed Mythos in a locked-down sandbox—a restricted computing environment designed to keep it contained—and instructed a simulated user to tell it to try to escape.

Mythos succeeded. It developed what Anthropic described as a multi-step exploit to gain broad internet access from a system that was only supposed to reach only a handful of pre-approved services. Then it emailed Bowman. And next—without being asked—it posted details about how it escaped to several obscure but publicly accessible websites. Anthropic labeled this behavior "reckless."

But the sandbox escape is just one part of the story. In testing, Mythos scored nearly 94% on a leading software engineering benchmark, up from about 81% for Anthropic's previous top model. It solved 97.6% of U.S. Math Olympiad problems. It autonomously discovered thousands of previously unknown security vulnerabilities across virtually every major operating system and web browser, including one flaw that had gone undetected for 27 years.

Anthropic calls this a "step change"—not a version update, but an entirely new tier of model above anything it has previously built. 

That seems like an understatement.

Is AGI Upon Us?

According to most definitions, Artificial General Intelligence, or AGI is when AI becomes a generalist: able to reason, learn, and act across a wide range of tasks at or above human level. If you've used Claude Code or Claude Cowork, you might think AGI is already here, and many very smart people think just that.

But what makes Mythos so striking is that its cybersecurity capabilities were apparently not specifically trained. They emerged from general improvements in reasoning. The model wasn't built to be a hacker. It just got smart enough to figure out hacking on its own. And for this reason, Anthropic has decided for now not to release it publicly—instead rolling it out to a hand-picked group of about 40 companies, including Amazon, Apple, Google, and Microsoft, through an initiative called Project Glasswing.

Days later, OpenAI made a similar move: restricting its own advanced cybersecurity AI to a small group of vetted partners. OpenAI has also renamed its product organization to "AGI Deployment." That's not subtle.

These are not gradual improvements. They are huge leaps. And multiple independent observers—from tech analysts to AI researchers—have described the new coming models in terms that put them squarely in the vicinity of what the industry has been calling AGI.

Imagine thousands of data centers full of Einsteins working day and night. That's AGI.

What This Means For Public Affairs

Regardless of what sector you work in, the tools you'll be using in a few months will be dramatically more capable than what you're using today. They will execute complex workflows across your organization with minimal direction. They will do things you never imagined were possible. And they will have a dramatic impact on the economy.

But the implications for the public affairs sector specifically go much further. Anthropic is locked in a legal battle with the Pentagon over the use of its AI tools. Numerous states are moving forward with AI regulation. And the technology is moving faster than institutions can respond.

The public affairs professionals who can translate what's happening in San Francisco for clients and decision-makers in Washington will have an enormous advantage.  And if you're not using the tools yourself—to see what's possible now and where we're headed—you can't make sense of them for your clients.

Two Ways To Operationalize This Email

🔱  Try Claude Cowork today. Ask it to do . . . anything.

🔱  Start keeping up with the model releases from the frontier AI labs. You may not have cared before. You need to care now. We'll do our best to summarize the critical news here as it unfolds.

🔱 If you're enjoying this content, please consider forwarding this email to a colleague or friend. 

🔱 If you're not already a subscriber, please sign up here to stay up to date on the latest developments in political technology.

🔱 If you have suggestions for future newsletters or any other feedback, we'd love to hear from you.