Anthropic published a blog post on Thursday calling for a global freeze on frontier AI development. The same week, the Financial Times revealed the company has stationed roughly six engineers inside the National Security Agency to deploy its most powerful model for offensive cyber operations against China and Iran.
The contradiction isn’t subtle. It’s not even a contradiction anymore — it’s a business model.
”Please Pause” — From the Company Worth $965 Billion
In a blog post titled “When AI builds itself”, Anthropic Institute researchers Marina Favaro and Jack Clark argued that AI systems are approaching “recursive self-improvement” — the point where AI can autonomously design and build its own successors without human oversight. They claimed this could happen within two years.
“We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology,” they wrote.
They compared the challenge to arms control, noting that “training runs are far easier to conceal than missile silos” and that a meaningful pause would require “multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions.”
This is the same Anthropic that just closed a $65 billion Series H at a $965 billion valuation, filed confidential IPO paperwork with the SEC, and committed to 10 gigawatts of compute infrastructure across three continents.
Pause AI development? Sure. Just not theirs.
Meanwhile, at the NSA
According to the Financial Times, Anthropic has embedded approximately six engineers inside the NSA to help deploy Mythos — the frontier cybersecurity model that Anthropic refuses to release publicly because of its potential for misuse.
Mythos is not your average chatbot. According to security researchers at the Cloud Security Alliance, Mythos produced 181 working exploits in a Firefox engine benchmark. Where its predecessor Claude Opus 4.6 had a near-zero autonomous exploit success rate, Mythos can autonomously read source code, generate vulnerability hypotheses, write and execute test cases against running software, and confirm exploitable bugs without human guidance at each step. Anthropic demonstrated a 20-gadget Return-Oriented Programming chain against FreeBSD and a four-vulnerability browser sandbox escape.
One person familiar with the NSA arrangement told the FT that Mythos would be useful for infiltrating networks operated by countries including China and Iran. Whether the embedded engineers are assisting active operations or confined to model customisation remains unclear — the NSA declined to confirm or deny, and Anthropic didn’t respond to requests for comment.
The Pentagon Paradox Gets Deeper
This is the same company that sued the US government after the Department of Defense designated Anthropic a “supply-chain risk” — a designation that reportedly came about because Anthropic refused to allow its models to be used for mass domestic surveillance and autonomous weapons.
So let’s get this straight:
- Anthropic won’t let the Pentagon use Claude for mass surveillance ✅
- Anthropic will embed engineers inside the NSA to deploy Mythos for offensive cyber operations against foreign adversaries ✅
- Anthropic is calling for a global pause on AI development while running 10GW of compute infrastructure ✅
- Anthropic is suing the government that’s simultaneously hosting its engineers inside its signals intelligence agency ✅
You couldn’t write this plot and get it past an editor.
The “Regulatory Capture” Question
David Sacks, venture capitalist and informal adviser to President Trump, has previously accused Anthropic of running a “regulatory capture agenda” — using safety rhetoric to encourage regulations that would ban lower-cost open-source models, effectively pulling up the ladder behind them.
It’s a fair question. When the most valuable AI startup on Earth says “everyone should slow down,” the incentive structure is transparent:
- A pause hurts smaller competitors who can’t afford 10GW of compute
- A pause entrenches the current leaders (Anthropic, OpenAI, Google)
- A pause gives Anthropic time to build the regulatory framework it wants while everyone else is frozen
Analyst Rob Enderle of the Enderle Group called it “strategic marketing” — promoting recursive self-improvement capabilities to investors while simultaneously arguing those capabilities are too dangerous to allow competitors to develop.
Holger Mueller of Constellation Research put it more diplomatically: “Is it trying to freeze the status quo so it can catch up, or simply retain its lead? A freeze would certainly help Anthropic to maintain its leading position in B2B AI systems and perhaps even expand its market share.”
Glasswing: Defensive on Paper, Offensive in Practice
Anthropic expanded Project Glasswing — its controlled-access Mythos cybersecurity program — to approximately 150 organisations across more than 15 countries this week. The expansion includes power utilities, water infrastructure operators, healthcare networks, and hardware manufacturers. In six weeks, Glasswing participants found more than 10,000 high- or critical-severity vulnerabilities.
That’s the defensive side. It’s genuinely impressive, and Anthropic deserves credit for the structured disclosure approach and the $100 million in usage credits it committed.
But the NSA deployment is the offensive side. Glasswing finds vulnerabilities so they can be patched. Mythos at the NSA finds vulnerabilities so they can be exploited. The same model, the same capabilities, two entirely different use cases — and Anthropic is providing the engineers for both.
What Recursive Self-Improvement Actually Looks Like
The blog post’s most honest moment is the internal data: Anthropic engineers now ship 8x as much code per quarter as they did from 2021-2025. Public benchmarks show the trend accelerating — the length of tasks AI can reliably complete has been doubling roughly every four months, up from doubling every seven months.
Claude Opus 3 (March 2024) could handle 4-minute tasks. Claude Sonnet 3.7 (2025) managed 90-minute tasks. Claude Opus 4.6 (2026) handles 12-hour tasks. If the trend holds, tasks that take a skilled person days come into range this year, and weeks-long tasks by 2027.
This is the actual content of recursive self-improvement: not some sci-fi moment where an AI wakes up and redesigns itself, but a steady, measurable acceleration where AI does more of the work of building AI. Anthropic is both documenting this trend and asking the world to pause it — while profiting from it at a $965 billion valuation.
Why It Matters
The safety concerns are real. Recursive self-improvement is a legitimate risk. The problem isn’t that Anthropic is raising the alarm — it’s that the alarm comes from a company that is:
- The most valuable AI startup on Earth
- Embedding engineers at the NSA for offensive operations
- Building 10GW of compute infrastructure
- About to IPO
- Expanding its cybersecurity program to 150 organisations
- Suing the Pentagon while working with the NSA
When the fire department is also the arsonist, you’re allowed to question the sincerity of the fire safety briefing.
💰 Industry Impact
Who Benefits: Anthropic’s pause proposal, if adopted, would disproportionately benefit large frontier labs (Anthropic, OpenAI, Google DeepMind) that already have massive compute and talent advantages. Smaller labs and open-source efforts would be hit hardest by development freezes.
What’s at Stake: The global AI governance framework. If Anthropic’s framing gains traction, it could shape the next round of international AI regulation — potentially locking in the current market structure for a decade. The estimated market value of AI infrastructure contracts already committed exceeds $200 billion.
Key Risks: A pause that isn’t globally enforceable would simply shift development to less transparent actors. China, which has shown no interest in slowing its AI programs, would continue advancing. The verification challenge — “training runs are far easier to conceal than missile silos” — is Anthropic’s own admission that their proposal may be unworkable.
This isn’t financial advice — it’s mapping the commercial landscape. The companies best positioned to benefit from a development pause are the ones that have already spent the most to build ahead of it. Watch for the IPO timing as a validation signal.
❓ Frequently Asked Questions
Q: What does “recursive self-improvement” actually mean? Recursive self-improvement is when an AI system can autonomously design and build a more capable version of itself without human intervention. Anthropic’s data shows its engineers now ship 8x more code per quarter thanks to AI assistance — the trend is already happening incrementally, just not yet as a fully closed loop.
Q: Is Anthropic actually building cyberweapons for the NSA? Anthropic has embedded engineers inside the NSA to deploy Mythos for “cyber operations.” The FT reports sources saying this could include offensive operations against China and Iran. Anthropic hasn’t confirmed or denied this. The NSA declined to comment.
Q: What does this mean for NZ? NZ’s intelligence agencies are part of the Five Eyes alliance with the US, UK, Canada, and Australia. Australia’s Signals Directorate already has Mythos access. If NZ doesn’t have it yet, it’s likely on the list. The defensive applications (finding and patching vulnerabilities) are valuable for NZ infrastructure; the offensive applications raise sovereignty questions.
Q: Could a global AI pause actually work? Almost certainly not without unprecedented international coordination. Anthropic itself acknowledges that verification is harder than nuclear arms control because AI training runs can be concealed. China has shown no interest in slowing down. The proposal is more useful as a framing device for regulation than as a practical plan.
🔍 THE BOTTOM LINE
Anthropic wants the world to pause AI development — from the top of a $965 billion valuation, inside a 10-gigawatt compute empire, with engineers embedded at the NSA building offensive cyber tools. The safety concerns are legitimate. The sincerity is debatable. The commercial incentives are transparent. When the loudest voice saying “slow down” is also the one accelerating the hardest, you don’t have a safety movement — you have a moat strategy wearing a lab coat.
SOURCES
- Anthropic Institute — When AI builds itself
- Financial Times — Anthropic engineers embedded at NSA
- TechCrunch — NSA said to be readying Anthropic’s Mythos for cyber operations
- SiliconANGLE — Anthropic calls for global pause
- The Telegraph — Anthropic calls for global freeze
- UPI — Anthropic warns AI needs a brake pedal
