Anthropic said on Friday it would suspend access to its two latest AI models, Claude Fable 5 and Mythos 5, after receiving an export‑control directive from the U.S. government. The letter, delivered at 5:21 p.m. ET, warned that the models could be exploited through a newly discovered “jailbreak” technique that might compromise national security. In response, the company cut off the models for every customer, foreign‑national employees included, and posted a detailed explanation on its blog.
The directive arrived amid a broader clash between Anthropic and the Trump administration. Earlier this year, the Department of Defense labeled the firm a "supply chain risk" after Anthropic pushed back against the military’s use of its technology. That designation barred federal agencies and contractors from employing Anthropic’s AI, prompting the company to file lawsuits challenging the ruling.
Claude Fable 5, a variant of Anthropic’s Mythos model, launched publicly with built‑in safeguards that block queries about cybersecurity, biology and chemistry. Anthropic claimed the rollout was coordinated with the U.S. government and followed a limited preview of Mythos in April, designed to let organizations test powerful cybersecurity features while limiting abuse. The company said the safeguards were meant to help firms bolster defenses without giving attackers a shortcut.
In its blog post, Anthropic explained that the government’s concern centered on a method that could bypass those safeguards. The company reviewed a demonstration where the technique identified a handful of known, relatively simple software vulnerabilities. Anthropic noted that other publicly available models can discover the same flaws without any jailbreak, suggesting the risk was not unique to its technology.
"Our understanding is that the government believes it has become aware of a method of bypassing, or ‘jailbreaking’ Fable 5," the blog read. "We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities. These vulnerabilities all appear relatively simple, and we have found that other publicly‑available models are able to discover them as well without requiring a bypass."
The firm argued that the identified jailbreak was narrow and would not make an attacker significantly more dangerous than using any other AI model. According to Anthropic, the only evidence the government provided was verbal, describing a scenario where a user asks the model to read a specific codebase and fix software flaws. The company said that one such potential jailbreak had been shared with officials.
White House and U.S. Commerce Department spokespeople did not immediately respond to requests for comment. Anthropic’s CEO, Dario Amodei, reiterated in a policy essay earlier this week that the company supports a transparent, structured government process for blocking unsafe AI releases. He criticized the current action as falling short of those principles, adding that the firm had already taken steps to mitigate misuse.
Anthropic’s decision to pull the models underscores the growing tension between rapid AI innovation and regulatory scrutiny. As lawmakers and agencies grapple with how to balance national security with technological advancement, companies like Anthropic find themselves navigating an increasingly complex compliance landscape.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.