Anthropic Is Selling Claude's Power While Warning About Its Dangers

At Code w/ Claude, Anthropic's London developer event, the energy was pure tech exuberance. Boris Cherny (who created Claude Code) talked about reconnecting with the magic of programming. Developers in the audience talked about a renaissance in coding. Anthropic demoed new features, new capabilities, and new ways to make Claude do more work faster. It was a masterclass in product marketing—and completely devoid of warnings about AI safety.

One moment stood out: An Anthropic researcher asked the crowd how many had shipped Claude-generated code without reading it first. A shocking number of hands went up. The researcher's response? "That's a dangerous game." Then he moved on to talking about how to squeeze more productivity from Claude Code.

Forty-eight hours later, Clark gave a lecture at Oxford that could have been from a different company entirely. He warned that AI posed a non-zero risk of killing humanity. He predicted that by 2028, AI would reach recursive self-improvement—the ability to improve itself without human intervention. He said most of the world is "in denial" about how capable current AI already is, let alone what's coming.

Clark's message was stark: We've underestimated how fast this is moving. When Claude Mythos (a model so powerful Anthropic didn't release it publicly) finished training, even Anthropic was shocked at how good it was. "We were like, 'Oh, it's here! It's here faster than we thought! And we've done insufficient preparation.'" That quote should keep everyone awake at night.

Here's what's actually happening: Anthropic is using developer enthusiasm to fund research that's making AI more capable at an alarming rate. That same money goes to funding research and policy work that openly warns about those same dangers. It's the opposite of hypocrisy—it's actually more honest than what other companies are doing. But it's also structurally unstable.

For enterprise buyers (which is you, if you're deploying Claude at scale), this duality actually matters. It means Anthropic is probably not bluffing about the risks—if they were, why would Clark be warning Oxford? It also means Anthropic is probably not going to slow down building capabilities just to be "safe." The company needs to ship fast to fund its research. You need to understand both sides of that tension when evaluating Claude for production workloads.

If you're building with Claude: Keep reading the code your AI generates. Anthropic's honest about the risks, which makes them more trustworthy than competitors who aren't. But that also means you can't assume Claude will always do the right thing. Guardrails, testing, and human review matter more now, not less.

If you're evaluating Claude for enterprise: The fact that Anthropic's founders are warning about AI safety is actually a selling point. They're not hiding the risks—they're being vocal about them publicly. That transparency is worth something.

TIME: "A Tale of Two Anthropics" by Billy Perrigo — The original reporting from Anthropic's London developer event and Jack Clark's Oxford lecture. Anthropic: Introducing Claude Mythos — Details on Project Glasswing and Claude Mythos's capabilities. Jack Clark's work on AI risk measurement — Clark's Alignment Research Center published research on measuring AI capabilities and risks.