Anthropic says it was forced to shut down access to its new Claude Fable 5 and Claude Mythos 5 AI models after the Trump administration issued a broad warning against use of the models by foreign nationals. The warning, from the Commerce Department, would have banned even many of Anthropic’s own employees from using or working on the models.
That dramatic result came after a multiday back-and-forth involving numerous administration officials and input from tech industry leaders. The effects of the government’s effectively shutting off access to one company’s AI models could have profound implications for future AI policy. What s is a timeline of the events that led to the effective ban on Anthropic’s Mythos-class models.
Thursday night (June 11)
Thursday evening: Amazon raises concerns with Trump administration officials
Amazon, Anthropic’s biggest investor, contacted senior White House and administration officials, saying its researchers had identified what they believed was a way to coerce, or “jailbreak,” the Fable 5 and Mythos 5 models into assisting users with risky cybersecurity questions. Anthropic released the models with safeguards designed to prevent responses to such queries. Axios reported that at least five other tech companies raised similar concerns with administration officials around the same time.
Late Thursday night: Administration begins treating the issue as urgent
White House and Commerce Department officials took Amazon’s concerns seriously, worrying that foreign bad actors might bypass Anthropic’s security guardrails and use the models to identify and exploit software vulnerabilities, Axios reported. The White House began trying to contact Anthropic, asking for an immediate reply.
Friday morning (June 12)
Friday morning: White House and Commerce officials engage Anthropic
Administration officials reportedly spent hours Friday trying to convince Anthropic to voluntarily shut off access to the Fable 5 and Mythos 5 models. It’s unclear from the reporting whether the government also asked Anthropic to fix the specific safety guardrail that allowed Amazon researchers to jailbreak the system, or to voluntarily restrict foreign access to the models.
Friday morning: Anthropic declines to voluntarily suspend the models
Axios reports that the government tried to persuade Anthropic to pause the release of the Fable 5 and Mythos 5 models, but was unsuccessful. Anthropic has said publicly that the government did not provide detailed technical evidence of the alleged issue and that the concerns involved only verbal descriptions.
“We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities,” Anthropic said in a Friday, June 12, blog post. “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.”
Friday afternoon
Around 1 p.m. ET: ‘You have 90 minutes to comply’
Anthropic received a call from administration officials around 1 p.m. ET, warning that unless the company acted, the government would move forward with export restrictions affecting the models, Reuters and Axios reported. Axios further reported that Anthropic was given roughly 90 minutes to comply.
5:21 p.m. ET: Commerce Department order arrives
At 5:21 p.m. ET Friday, the Commerce Department sent an enforcement order restricting access to Fable 5 and Mythos 5 by foreign nationals, citing national security concerns.
Friday evening and night
Friday evening: Anthropic disables Fable 5 and Mythos 5
Anthropic said it disabled access globally because the order restricted access by foreign nationals and the company could not immediately separate eligible from ineligible users. It erred on the side of caution and compliance.
Context
About Fable 5 and Mythos 5
Fable 5 and Mythos 5 are the first publicly released models derived from Anthropic’s Mythos family. The original Mythos model showed unusual skill during training at finding software bugs and exploiting them to disrupt or take control of systems. That raised enough concern inside Anthropic that the company grouped cybersecurity with other high-risk domains, including biology and chemistry, when setting limits on Mythos-derived public models. For Fable 5 and Mythos 5, that means prompts flagged as sensitive in those areas are routed to Claude Opus 4.8, a less capable model with its own guardrails.
About the politics
Many observers saw the tech industry’s surprising support for Donald Trump’s presidential candidacy in 2024 as part of an implicit bargain over AI policy. Trump pledged that his administration would allow the AI industry to regulate itself and do what it could to advance the tech industry’s goal of preventing states from passing their own AI laws in the absence of federal legislation. The Trump administration has also ordered federal agencies to drop numerous investigations and enforcement actions involving big tech companies. So the administration’s recent concerns about the national security risks of the largest AI models, and its pivot toward involvement in safeguarding models Mythos, is striking. (See: Trump’s June 2026 executive order.)
The current dustup might also be partisan and political. In March, Defense Secretary Pete Hegseth clashed with Anthropic after the AI company refused to allow its AI to be used for autonomous weapon targeting or domestic surveillance. The Pentagon then designated Anthropic a “supply-chain risk,” effectively blacklisting it from much federal defense work. Anthropic sued.
Security fears
Before launch, Anthropic said internal and external red teams spent more than 1,000 hours trying to fool Fable 5 into dispensing information on banned topic areas. Anthropic said the teams had no success identifying universal jailbreaks.
However, at least one independent AI researcher appears to have found a successful jailbreak before Amazon researchers went to the government with their report. The researcher, using the handle Pliny the Liberator, claimed on X to bypass Fable 5’s filters using a multi-agent approach involving a previously jailbroken Claude Opus 4.8 model, along with techniques including query decomposition, long-context framing, fiction and narrative structures, and academic taxonomies.
Sumber Artikel:
Fastcompany.com
