Anthropic safety warnings draw scrutiny after US export ban
A Financial Times analysis found Anthropic used far more risk language than OpenAI before Washington restricted foreign access to two models.
By Maya Lindqvist · Senior Technology Correspondent
4 min read
Washington’s decision to bar foreign nationals from using Anthropic’s newest AI models has put the company’s safety messaging under fresh scrutiny. The move matters beyond one company because executives and officials in Europe and Silicon Valley fear it could signal tighter US controls on access to frontier AI systems.
The Financial Times reported that Anthropic has spoken about AI risks, regulation and restrictions far more often this year than OpenAI. In an analysis of official statements, social media posts and articles by Anthropic and chief executive Dario Amodei, the FT found that five words in every 1,000 related to those themes. For OpenAI and Sam Altman, the comparable rate was 0.6 words per 1,000.
The comparison has become politically sensitive after the US government last week blocked foreign nationals from using Anthropic’s latest models, Mythos and Fable. The FT reported that some technologists have linked the decision to Anthropic’s repeated warnings about the risks of advanced AI, particularly around Mythos.
Yann LeCun, Meta’s former chief AI scientist, criticized Amodei on social media, saying the ban showed that Amodei’s “ridiculous fear-mongering” had paid off. “One reaps what one sows,” LeCun wrote, according to the FT.
FT analysis finds heavier risk language
The FT said it built lists of terms such as “harmful,” “dangerous” and “misaligned,” then measured their frequency in company and executive communications. It also used sentiment analysis to compare the tone of Anthropic’s and OpenAI’s public messaging.
In Anthropic communications this year, the FT found 336 uses of “risk,” 121 uses of “safeguard” and 128 uses of “vulnerability.” OpenAI used the same words 30, 33 and 10 times respectively, according to the analysis.
Altman appeared to criticize the marketing around Mythos in an April podcast interview. Asked about the model, he said: “It is clearly incredible marketing to say, ‘We have built a bomb. We are about to drop it on your head.’”
Anthropic has long presented itself as a safety-focused AI company. The FT reported that it often publishes research and public statements on possible harms from AI and has called for more government intervention.
Days before the export ban, Amodei wrote on his personal website that regulators were moving too slowly. He said Mythos showed “very real risks to cyber security,” including possible disruption to finance, critical infrastructure and national security.
The FT said that post was not the most negative Anthropic publication it reviewed. An April 2025 item on rare AI behaviors, including sabotage and weapons-related information, used about three times more negative language.
The analysis also found that Anthropic’s language has become less severe since 2023 as its products gained users. Its use of terms tied to risk and regulation has roughly halved from the same period in 2023, the FT reported.
Industry and political fallout
Anthropic has described Mythos as able to find critical cyber security flaws and initially limited access to certain US organizations on safety grounds. The company had been working with government officials on a controlled rollout before a wider release earlier this month, according to the FT.
Data from AlphaSense showed that Mythos drew far more media attention than other AI models released this year, the FT reported. Mentions rose after Anthropic announced the model in April and rose again after the export ban.
David Sacks, a former US government AI tsar, wrote on X that a “credible trusted partner” had warned the administration about a way to bypass guardrails on Fable. He said Anthropic downplayed the concern, prompting the government to impose the ban reluctantly, according to the FT.
The ban follows public disputes between Anthropic and senior US officials over domestic surveillance and lethal autonomous weapons. The Pentagon in February named Anthropic a supply-chain risk to national security, and the two sides are in litigation over that designation, the FT reported. Anthropic declined to comment.
YouGov polling cited by the FT found that a majority of respondents supported effective regulation even if it slowed technological progress. French President Emmanuel Macron said this week the dispute had “clarified the stakes” for the US and its G7 allies, and called for stronger AI regulation while warning against “non-cooperation among democracies.”
Lennart Heim, an independent AI policy researcher and former Rand think-tank staffer, told the FT the US response did “not inspire confidence,” citing the administration’s pro-innovation stance, its support for exporting advanced AI chips to China and its criticism of safety-focused regulation.
This story draws on original reporting from Ars Technica.