US Bars Foreign Nationals From Anthropic Fable 5, Mythos 5 as Colgate AI Hits 90%

Crypto News

The United States Commerce Department moved on June 12 to suspend Anthropic from offering its newest flagship models, Claude Fable 5 and Claude Mythos 5, to any foreign nationals, invoking national security authority. The restriction reaches beyond overseas users: it explicitly covers foreign-national Anthropic employees, including those working inside the United States. The company's official statement confirmed the scope with the phrase “including foreign national Anthropic employees.” The order creates an unusual situation in which some of the researchers who helped build the models could, in theory, lose access to them. Anthropic declined to specify which staff would be affected.

The trigger, according to the company, was a dispute over so-called jailbreak capability. Anthropic says a competitor demonstrated to the government that Mythos 5 could read entire codebases and patch software vulnerabilities, raising national-security concerns. The company pushed back, arguing the model merely identifies a small number of already-known minor flaws and that other publicly available systems, including GPT-5.5, offer comparable abilities. The episode underscores how quickly frontier-model capabilities around autonomous code analysis have become a regulatory flashpoint. It also illustrates the difficulty of drawing clear lines between routine developer tooling and dual-use functionality that governments may classify as sensitive.

The timing drew particular attention. Anthropic chief executive Dario Amodei had on June 10 publicly called for stronger oversight of advanced AI, arguing that governments should hold the power to block deployment of high-risk models. Two days later, the Commerce Department applied a similar authority against Amodei's own flagship systems. The reversal prompted public commentary from Turing Award winner Yann LeCun, who wrote that the company's alarmist messaging had finally caught up with it and added the phrase “one reaps what one sows.” The exchange highlights a widening rift among senior AI figures over how aggressively the sector should be regulated.

Elsewhere in the sector, consumer-goods giant Colgate-Palmolive and research group PyMC Labs published findings suggesting large language models can act as “synthetic consumers” in product testing. Drawing on 57 personal-care product concept tests and data from 9,300 real United States consumers, the study reported that AI-simulated purchase-intent rankings matched up to 90% of human test-retest reliability. The work, covering toothpaste, oral and personal-care categories, indicates brands could one day screen new concepts through thousands of AI simulations before committing to expensive human surveys. For large firms that means faster, cheaper iteration; for smaller brands it could open access to consumer insight previously reserved for major budgets.

The researchers found that asking a model directly whether it would buy a product produced distorted, overly cautious answers clustered around middle values. Instead, they assigned the model a detailed consumer persona defined by age, income, region and demographic background, then let it react in natural language before converting that response into a one-to-five score using a method called Semantic Similarity Rating. The technique reproduced realistic patterns, including lower purchase intent among lower-income personas and higher willingness to try new products among middle and higher earners. Stripping out the persona, however, made the model uniformly optimistic, collapsing its ability to rank products accurately.

In Taiwan, payment processor ECPay released an open API Skill package on GitHub, now at version 3.2, that converts its payments, logistics, e-invoicing and ticketing documentation into a knowledge base readable by AI coding tools. Developers can plug the Skill into OpenAI Codex CLI, Google Gemini CLI, ChatGPT GPTs, Claude Code, Cursor and GitHub Copilot CLI, then ask the assistant to generate integration code, debug errors or switch between test and production environments across twelve supported languages. The move mirrors a broader trend of financial-infrastructure providers, including those building blockchain rails, packaging their systems for AI-assisted development.

Taken together, these developments mark a turning point in how AI capability, regulation and commercial deployment intersect — a theme that increasingly overlaps with digital-asset markets, where automated trading and on-chain tooling lean on the same frontier models now facing export limits. COINOTAG's aggregate market data frames the backdrop: the Fear and Greed Index sits at 13, deep in extreme fear, while Bitcoin dominance stands at 70.4% and total altcoin-inclusive market capitalization holds near $1.83 trillion. That risk-off posture, consistent with bear-market conditions, suggests investors are weighing regulatory tail-risk across both AI and DeFi infrastructure as the two sectors converge.