Headline: Anthropic engineers embedded at NSA to deploy Mythos — and the company is calling for a global AI pause
Anthropic has quietly placed roughly six engineers inside the U.S. National Security Agency to help customize its most capable model, Mythos, for offensive cyber operations, the Financial Times reports. These forward-deployed staff are adapting the model for specific tasks; one source told the FT that Mythos “could be useful” for infiltrating foreign networks, naming countries such as China and Iran. It’s not confirmed whether the engineers are directly participating in live operations.
Key facts
- Mythos is Anthropic’s most powerful model and one the company has declined to release publicly because of misuse risk. Access is limited to vetted partners via Project Glasswing — a restricted consortium that includes Microsoft, Apple and Amazon.
- Anthropic is entangled in a legal and political fight with the U.S. military. In late February, Defense Secretary Pete Hegseth labeled the company a supply-chain risk after a $200 million contract collapsed. The sticking point: Anthropic refused Pentagon requests to allow Claude to be used for fully autonomous weapons or domestic mass surveillance. The NSA contract was exempt from that restriction.
- A California judge blocked the DoD blacklisting as possible First Amendment retaliation; a D.C. appeals court denied Anthropic’s bid to pause the listing while litigation continues. According to the FT, the NSA continued using Mythos throughout this period.
Why Anthropic is worried about what AI can build
On the same day the NSA story broke, the Anthropic Institute published “When AI Builds Itself,” a paper describing how Claude has increasingly automated its own development and why that trend is cause for alarm.
Important data points from the report:
- Since Claude Code launched in early 2025, Claude now writes more than 80% of the code merged into Anthropic’s production codebase — up from low single digits.
- Engineers at Anthropic ship about eight times as much code per day as they did in 2024.
- In April, Claude agents were tasked with an open AI-safety research problem (whether a weaker model can reliably supervise a stronger one). Humans recovered about 23% of the performance gap in roughly a week. The agents themselves recovered 97% after roughly 800 cumulative compute hours — with humans only setting the question while the agents designed and ran every experiment. Anthropic calls this the first published evidence of Claude exercising research judgment rather than merely executing human-specified tasks.
The company’s central fear is “recursive self-improvement”: if AI systems autonomously design, build and train their successors, humans could be sidelined from crucial control points. Small misalignments today could compound across generations of self-improving systems, the paper warns. Anthropic’s proposed remedy is a verifiable global pause — multiple frontier labs suspending certain lines of development simultaneously, with independent checks to ensure compliance. Anthropic says it would join such a pause but admits a unilateral slowdown only hands the lead to whoever refuses to stop.
Context and stakes
This is not new rhetoric from labs that profit heavily from AI: the same organizations that warn of existential risk have powerful commercial incentives to keep developing. In 2023, numerous AI researchers signed public appeals for global mitigation efforts; despite that, major labs continued advancing models. Anthropic itself has continued development while simultaneously sounding the alarm.
Why crypto audiences should care
- Nation-state cyber tools that leverage advanced LLMs could accelerate or sophisticate attacks on exchanges, custodians and DeFi infrastructure — areas where operational security and key management are critical.
- More powerful AI for offensive operations raises the bar for defenders: automated discovery, exploit development, and social engineering could scale faster than current defenses.
- The same model-driven automation Anthropic describes (AI writing, testing and shipping code) could change how smart contracts are audited, how wallets and custodial systems are developed and how vulnerabilities are found — for better and for worse.
- The public legal fight and the looming August deadline for the Pentagon to drop Claude from its systems intersect with Anthropic’s expected IPO, a moment that will make the company’s finances and governance more transparent to markets.
What to watch next
- How courts resolve Anthropic’s litigation with the Pentagon and whether the supply-chain designation sticks.
- Any official disclosures about the NSA program’s scope and whether Mythos has been used in active operations.
- Whether a credible, independently verifiable pause proposal gains industry buy-in — and what verification mechanisms would look like.
- The broader impact on crypto security posture as AI-driven offensive tools and automated development workflows mature.
Bottom line: Anthropic’s engineers are reportedly embedded at the NSA to tailor an unusually capable, tightly guarded model for offensive cyber use. At the same time the company is publicly warning that AI is on a trajectory toward self-directed development and is advocating for a verified global pause — a tension between building powerful tools and trying to stop the race to the next frontier.
Read more AI-generated news on: undefined/news
