NSA Deploys Anthropic’s Mythos AI vs Pentagon — 20 Apr 2026

America’s signals intelligence agency is quietly running Anthropic’s most powerful restricted model on classified networks, despite an active Pentagon legal fight. Adobe data shows AI agents have flipped the economics of online retail in a single year, Coinbase is building virtual teammates modelled on its most celebrated alumni, and a developer security breach tied to an AI tool has sent the crypto industry scrambling.

The NSA is running Anthropic’s Claude Mythos Preview on classified networks, even as the Department of Defense fights Anthropic in federal court. Sources cited by Axios confirmed the deployment today, creating an unusual split inside the US federal government. The DoD designated Anthropic a supply-chain risk in March and launched litigation to limit the company’s government contracts. An administration source told Axios that every federal agency except the DoD now wants access to Anthropic’s tools.

Claude Mythos is Anthropic’s most capable and most restricted model, made available only to a handful of vetted organisations since its launch earlier this month. The fact it has reached NSA classified networks suggests its capabilities are of significant intelligence interest. Anthropic CEO Dario Amodei met White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent on 17 April, with both sides describing the discussions as productive. Anthropic finds itself in a peculiar position: its most powerful model is running inside one of the world’s most secretive intelligence agencies, while the department that agency reports to is still trying to block it.

AI-driven traffic to US retail websites grew 393% year-on-year in the first quarter of 2026, and agentic shoppers now outperform human visitors on every commercial measure. The data comes from Adobe Analytics, which monitors more than one trillion visits to US retail sites and supplemented its findings with a survey of more than 5,000 US consumers. The shift in a single year is striking. In March 2025, AI referral traffic converted 38% worse than traditional channels such as paid search and email. By March 2026, the position had fully reversed: AI traffic now converts 42% better, a new high. Revenue per visit from AI referrals runs 37% above non-AI traffic.

Shoppers arriving from AI assistants spend 48% more time on page, view 13% more pages per visit, and show a 12% higher engagement rate than other visitors. Adobe Digital Insights director Vivek Pandya noted the data demonstrates a rapid maturation of AI as a commerce channel. A structural gap is emerging at the same time: Adobe’s AI Content Visibility Checker found that retail homepages score an average of 75% for readability by AI models, while individual product pages fall to just 66%. Retailers whose sites are not structured for AI agents risk becoming invisible to the next wave of buyer. McKinsey projects agentic commerce could drive $1 trillion in US retail revenue by 2030.

Coinbase is piloting AI agents built to replicate the strategic thinking of two of its most celebrated former executives, with CEO Brian Armstrong predicting that agents will eventually outnumber human employees. Armstrong announced on Saturday that the exchange has built agents modelled on co-founder Fred Ehrsam and former chief technology officer Balaji Srinivasan, both described as “legendary former Coinbase employees.” The agents live inside Slack and email as virtual teammates any staff member can consult. The Fred agent helps employees refine documents, strategies, and concepts. The Balaji agent is designed to challenge assumptions and spark innovation.

Engineer Travis Bloom reported consulting the Balaji agent about a new idea and said it “really helped crystallise the vision.” Armstrong said the next phase will allow any employee to spin up additional agent colleagues, though future versions will likely not be modelled on specific individuals. The announcement sits within a broader AI strategy at Coinbase. The exchange launched Agentic Wallets in February, allowing AI systems to hold funds and execute on-chain transactions without human approval, and is developing the x402 protocol as a payment standard for autonomous AI commerce. Armstrong’s conclusion: “I suspect we will have more agents than human employees at some point soon.”

A security breach at Vercel, linked to a compromised AI tool, exposed API credentials used by cryptocurrency application developers and prompted an industry-wide scramble to rotate keys. Vercel provides hosting infrastructure for many Web3 application frontends, the interfaces that connect users’ digital wallets to blockchain services. A breach at this layer can cascade through multiple downstream applications simultaneously. The incident points to a risk that security teams are still calibrating: AI tools integrated into developer workflows typically hold access to environment variables, authentication tokens, and API keys. A single compromised assistant can therefore expose credentials across an entire project stack. No specific AI tool had been publicly named at the time of writing, but developers were advised to audit all credentials, rotate any keys in environments where AI tools have been active, and review access logs for anomalous activity.

Here is everything else worth knowing from today’s AI news.

• Anthropic launches Claude Design — Anthropic Labs released Claude Design on 17 April, a new AI canvas product aimed at helping designers and product teams build interfaces and assets directly with AI. Anthropic
• Claude Opus 4.7 now available — Released on 16 April, Opus 4.7 supports images up to 3.75 megapixels (more than 3x previous resolution), improves long-horizon coding, and scores state-of-the-art on finance and legal reasoning benchmarks. Pricing is unchanged at $5 per million input tokens. Anthropic
• FCA consults on UK crypto regime — The Financial Conduct Authority launched a consultation on 15 April covering its incoming regulatory framework for cryptocurrency firms, including custody, trading venues, and stablecoins. Relevant as the regulatory perimeter for AI-adjacent crypto products continues to take shape. FCA
• US court blocks AI agents from buying on Amazon — A San Francisco judge issued a preliminary injunction in March preventing Perplexity’s Comet browser from completing purchases on Amazon without explicit user authorisation, setting an early precedent for agentic commerce law. Decrypt
• Adobe: retail product pages poorly structured for AI agents — Adobe’s AI Content Visibility Checker found that individual product pages score just 66% on average for AI readability, versus 75% for homepages. The best-performing retailers reach 82.5%; the lowest score 54.2%. Decrypt
• Coinbase Agentic Wallets and x402 protocol — Coinbase launched Agentic Wallets in February, enabling AI systems to hold and transact cryptocurrency autonomously. Its x402 protocol is designed to become the standard payment layer for AI agents transacting on-chain. Decrypt

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence — published every evening.