Anthropic's Safety Rhetoric May Have Fed Its Own Export Ban

Anthropic's Fable 5 and Mythos 5 models are now off-limits to users outside the United States after President Trump ordered the company to block international access, citing national security concerns. The proximate trigger, according to reporting including from Fortune, was a reported jailbreak vulnerability that raised alarms about whether the models could be exploited to identify software weaknesses. But a Financial Times analysis published this week makes a more uncomfortable argument: that Anthropic's own public communications may have helped frame the case against itself.

The FT's approach was quantitative. In Anthropic's 2026 official statements, "risk" appeared 336 times, "safeguard" 121 times, and "vulnerability" 128 times, versus 30, 33, and 10 times respectively at OpenAI. Roughly five in every 1,000 words Anthropic used in 2026 touched on risk, regulation, or restrictions, compared to 0.6 per 1,000 for OpenAI. The language Anthropic deployed to signal responsibility reads, in the FT's framing, as a dossier of national-security concern.

The timing made things worse. Days before the export ban, Amodei published a blog post on his personal website arguing that frontier AI models like Mythos posed "emblematic" threats to national security and that regulators were moving too slowly. Yann LeCun, Meta's former chief AI scientist, later said the ban showed that Amodei's "ridiculous fear-mongering" about AI had finally paid off.

The honest caveats are twofold. The FT analysis itself found that Anthropic had softened its risk language significantly since 2023, with usage of risk- and regulation-related terms roughly halving from that period. And correlation between a lab's communications and a government order is not causation: the government has not shared specific national-security reasons with Anthropic, and there was already a separate record of friction, including the Pentagon naming Anthropic a supply-chain risk to national security in February and public clashes over the company's technology in domestic surveillance and autonomous weapons. Whether the language drove the decision or merely provided post-hoc framing is a question the reporting does not settle.

For other AI labs, the tension is now explicit. Safety-forward language has served as a credibility signal for investors, researchers, and aligned policymakers. It may now also function as a liability in a policy environment where the same terms can be read as evidence of danger. The more consequential open question is whether labs that recalibrate their public vocabulary will do so in ways that affect actual safety practices, or only the language around them.