The Department of Commerce issued an export‑control directive late Friday, compelling Anthropic to cut off access to its two flagship models, Claude Fable 5 and Claude Mythos 5, for all users worldwide. The order arrived at 5:21 p.m. ET and was framed as a national‑security measure after officials warned that the models could be exploited to uncover software vulnerabilities.

Anthropic responded on its X account, confirming the shutdown and expressing frustration. In a detailed blog post, the company said it received only verbal evidence of a “potential narrow, non‑universal jailbreak” – essentially a prompt that coaxed the model into scanning a codebase for flaws. Anthropic maintains that this capability already exists in other publicly available models, including OpenAI’s GPT‑5.5, and is routinely used by cybersecurity professionals.

Claude Mythos 5, the company’s most capable model, debuted in early April under a tightly controlled rollout. Anthropic described Mythos as uniquely adept at identifying security holes across major operating systems and browsers. To prevent misuse, it limited Mythos to a select group of roughly 50 vetted organizations through Project Glasswing, a program that includes Amazon, Apple, Google, Microsoft and CrowdStrike. These partners use the model for defensive cybersecurity work.

Claude Fable 5, launched just three days before the government’s order, is a commercial‑grade version of Mythos with additional guardrails that block high‑risk topics such as advanced cybersecurity techniques and biological research. Benchmark tests from Vals AI, a firm that tracks AI performance, listed Fable 5 as the most capable publicly accessible model at the time of its release.

The export‑control action technically targets foreign‑national access, but Anthropic’s compliance extends to every user, effectively pulling the models from the global market. The company argues that the measure is disproportionate. “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” the blog read. Anthropic warns that applying the same standard across the industry would stall the deployment of frontier models from all providers.

Safety has been a cornerstone of Anthropic’s public image as it prepares for a potential IPO later this year. The company’s cautious stance on Mythos, branding it as a “model so dangerous it couldn’t be released publicly,” now appears to have backfired, drawing heightened scrutiny from regulators. Sam Altman of OpenAI previously called Anthropic’s marketing around Mythos “fear‑based,” noting the company’s strategy of selling a “bomb shelter” for its technology.

Anthropic’s internal safeguards rely on independent classifier systems that operate separately from the language model itself. According to the firm, even if a user manages to coax the model into ignoring a refusal, these classifiers continue to block dangerous outputs. The government’s directive suggests that officials are not convinced those layers are sufficient.

While the shutdown impacts only the two newest models, the broader implications could ripple through the AI industry. Companies may reconsider how they balance rapid model releases with regulatory compliance, especially as U.S. authorities tighten export‑control policies on advanced AI. Anthropic’s next steps remain unclear, but the company’s leadership has signaled it will continue to push for a more nuanced assessment of the alleged jailbreak.

Este artigo foi escrito com a assistência de IA.
News Factory APP - notícias agênticas para impulsionar seu SEO e AEO.