In short
- Anthropic says it has recruited about a half-dozen engineers at the NSA to use its Mythos AI model for offensive cyber operations — possibly including cyberattacks on China and Iran.
- Anthropic also warned that AI is on the verge of reinventing itself and calling for a global communication system.
- All of them came as Anthropic files for an IPO that could be worth more than $1 trillion.
Anthropic has placed about six engineers within the National Security Agency to help deploy Mythos — its most advanced AI model — on offensive cyber operations, The Financial Times reported Thursday.
Engineers are forward-deployed workers, depending on the type of work to be done. Another source told the FT that it could be useful for entering countries like China and Iran.
It has not been confirmed whether these engineers are involved in permanent projects. What: Legend is the same brand Anthropic has refused to release publicly, citing the risk of misuse. The company gave money to those who voted Project Glasswing-a restricted partnership that includes Microsoft, Apple, and Amazon.
Anthropic is also suing the Pentagon. At the end of February, Defense Secretary Pete Hegseth designated the company as vulnerable to auction – a mark reserved for foreign rivals like Huawei – after a $200 million deal fell through. The sticking point: Anthropic refused to let the DoD use Claude for autonomous weapons or mass surveillance. The NSA contract was exempted from that ban.
California judge canceled registration as the first visible revenge of the Revolution. A D.C. appeals court denied Anthropic’s motion to stay the proceedings pending. The NSA still used Mythos regularly, the FT reports.
How to stop the AI that creates the AI
On the same day the NSA story broke, the Anthropic internal investigation was published “When AI Builds,” Let’s see how Claude arrived at his development. In it, the company argues against a global freeze in the AI arms race—and even compared it to the Cold War nuclear deal between the United States and Russia.
To help you understand why, the company provided the following:
Claude now writes more than 80% of the code included in Anthropic’s codebase—from the small number of Claude Code pre-implementation in early 2025. Engineers contribute about eight times as much code per day as they did in 2024.
The authors of the report – the Anthropic Institute is led by Marina Favaro and co-founder Jack Clark – say that this approach leads to what they call iterative self-improvement: AI systems that create, design, and train their successors, with humans playing a smaller role at each step.
In a visual presentation, researchers show the time when the first method of using AI at work as people makes computers to get results, and the increase in artificial intelligence machines AI Agents that act on subagents until the results are achieved, no people are involved.
The sharpest point he mentions: In April, Claude’s assistants were given the open-ended problem of AI security – whether a smaller version could manage a more powerful one – and left to run it. Two human researchers in about a week also found 23% of the difference in performance between samples. Agents recovered 97%, over 800 hours of testing. People ask a question. The facilitators performed each experiment. This is the first published case of Claude making an investigative decision, not just doing what someone else has mentioned.
That’s the line Anthropic is worried about crossing. When AI chooses the right experiment to run—rather than just run it—humans lose the last stage of civilization. Small flaws seen in today’s models can compound over generations until no one can fix them.
What they want is a guaranteed break in the world – several laboratories stopping at the same time, and making sure that everyone stops. Anthropic said it will join them. The gradual decline, he admits, only serves as a guide for anyone going.
We have seen this movie. The labs building AI are the same ones warning about the dangers of AI. However, AI is the most profitable business of the decade, so no one wants to stop it, even those who warn about AI.
Return in 2023Over a hundred big names in the AI research community have signed on an open letter asking for a global effort to reduce the risk of extinction that AI development has. A few months before it happensAnother open letter called for OpenAI to stop ahead of ChatGPT because of its risks.
No one stopped after the 2023 open letter. OpenAI didn’t. Anthropic did not. The Pentagon’s deadline to remove Claude from its operations falls in August, the same time Anthropic’s IPO is expected to move its funds to the public.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
