Anthropic dropped a bombshell this week that should've made every AI researcher in the world do a double-take. The company—famous for its Claude chatbot and its constant hand-wringing about existential risk—just floated the idea of a worldwide "temporary pause" on AI development while simultaneously being exposed for embedding engineers inside the National Security Agency to help with offensive cybersecurity operations targeting Iran and China. If this sounds like cognitive dissonance, you're not alone in noticing.

The Self-Improvement Problem

In a lengthy post published Thursday, Anthropic detailed how Claude is trending toward what researchers call "recursive self-improvement"—the ability for an AI system to design and build more powerful versions of itself. This capability sits at the center of doomsday scenarios that imagine AI agents iterating endlessly until they surpass human control entirely. Anthropic acknowledged this trajectory could lead to "humans losing control over AI systems." The company proposed convening policymakers, researchers, and civil society groups to answer these thorny questions—a conversation Anthropic has been chasing since its founding.

The NSA Connection

Here's where things get interesting. Alongside the safety manifesto, the Financial Times reported that Anthropic has engineers physically stationed at the NSA, helping the intelligence agency deploy the company's Mythos model for offensive cyber operations. This comes despite an ongoing legal battle with the Pentagon over how its tools are being used. Professor Steven Murdoch from University College London didn't sugarcoat his assessment: "Anthropic might give the impression of being warm and fuzzy, but their definition of AI safety is narrow. Supporting US authorities in the development of offensive capabilities has never been something they have spoken against."

Code Generation Stats and IPO Ambitions

The timing of this public relations offensive becomes clearer when you look at the numbers. Anthropic reported that as of May 2026, more than 80% of code merged into their codebase was authored by Claude—meaning the AI is already heavily involved in building its own infrastructure. The company also filed for an IPO this week with a rumored valuation hitting $1 trillion. Heidy Khlaaf, chief AI scientist at the AI Now Institute, offered a blunt interpretation: she called Anthropic's Mythos announcement "a marketing post."

Key Takeaways

  • Anthropic wants a worldwide pause on AI development while actively supporting NSA offensive cyber operations against Iran and China
  • More than 80% of Anthropic's codebase is now AI-generated as of May 2026
  • The company filed for IPO with potential $1 trillion valuation this week
  • Experts question whether the safety rhetoric masks commercial interests

The Bottom Line

Anthropic wants to slow everyone else down while they sprint toward AGI and a massive public offering. Call me cynical, but when your definition of "AI safety" conveniently excludes offensive military applications, maybe the real risk isn't the technology—it's who controls it.