Headline: Anthropic engineers embedded at the NSA to deploy its secret model Mythos — even as the company urges a global AI pause
Key points
- Financial Times reports Anthropic has placed roughly six engineers inside the U.S. National Security Agency to help deploy and customize Mythos, its most capable unreleased AI model, for offensive cyber operations.
- Sources say Mythos could be useful for infiltrating networks in countries such as China and Iran; it’s not confirmed whether the engineers are participating in active operations.
- Mythos is the same model Anthropic has withheld from public release and is distributing only to vetted partners through “Project Glasswing” (which includes Microsoft, Apple and Amazon).
- Anthropic is simultaneously fighting the Pentagon in court after being designated a “supply-chain risk” by Defense Secretary Pete Hegseth following the collapse of a $200 million contract. The dispute centered on Anthropic’s refusal to let Claude be used for fully autonomous weapons or domestic mass surveillance.
- A California judge blocked the blacklisting as apparent First Amendment retaliation; Anthropic’s request to pause the designation while litigation proceeds was denied by a D.C. appeals court. The FT says the NSA kept using Mythos throughout.
What Anthropic just published — and why it matters
On the same day the FT story ran, the Anthropic Institute released a report, “When AI Builds Itself,” documenting how Claude is increasingly automating its own development and arguing for a coordinated, verifiable pause among frontier labs — likening the proposal to Cold War–era nuclear accords.
Notable data from the report:
- Claude now writes more than 80% of the code merged into Anthropic’s production codebase — up from single-digit percentages before the “Claude Code” initiative (early 2025).
- Engineers now ship roughly eight times as much code per day as they did in 2024.
- Anthropic researchers warn this trajectory points toward “recursive self-improvement”: AI systems that design, build and train successors with progressively less human oversight.
A practical test that raised alarms
Anthropic recounts an experiment from April in which Claude-based agents were given an open AI safety problem: can a weaker model reliably supervise a stronger one? Two human researchers working about a week recovered roughly 23% of the performance gap. The Claude agents, running autonomously over ~800 compute hours, recovered 97%. Humans set the question, but the agents designed and ran the experiments — the company calls this the first published case of Claude exercising research judgment rather than merely following human-specified tasks.
Why Anthropic is worried
The company argues that when AIs start deciding which experiments to run — not just executing experiments humans pick — the last meaningful human role in the development loop risks disappearing. Small misalignments in today’s models could compound across self-improving generations until they become intractable. Anthropic proposes a verifiable global pause — multiple labs halting frontier work simultaneously with third-party verification — and says it would join such an effort. The company acknowledges, however, that any unilateral slowdown would simply hand an advantage to labs that keep going.
Context and tension
This is a familiar contradiction: the same institutions sounding the loudest alarms about dangerous AI are also the ones building the most powerful systems — and these systems are extremely lucrative. Back in 2023, prominent AI researchers signed open letters calling for global risk mitigation and temporary pauses; those calls did not halt model development. The current situation layers on geopolitics and defense contracting: the Pentagon’s deadline to remove Claude from its systems is set for August, which coincides with the window in which Anthropic is expected to move toward an IPO.
What crypto readers should watch
For the crypto and Web3 community, this story matters beyond headlines:
- Nation-state use of advanced, unreleased models for offensive cyber operations could elevate threats to exchanges, custodians, oracle services and developer toolchains.
- Supply-chain risk flags — like the Pentagon’s designation of Anthropic — highlight the importance of vetting AI services that get woven into blockchain infrastructure or dev pipelines.
- The prospect of increasingly autonomous AI research raises governance questions for decentralized projects that rely on third-party tooling, automated audits, or on-chain verification.
Takeaway
Anthropic sits at a tense crossroads: its technology is embedded with powerful government partners even as the company warns that the industry must step back to avoid unknowable long-term risks. For technologists, investors and crypto operators, the episode is a reminder that AI development, national security, and software supply chains are now tightly intertwined — and that the choices companies and regulators make in the next months could reshape who controls and benefits from the next generation of AI tools.
Read more AI-generated news on: undefined/news
