Anthropic announced on Friday that it will immediately shut down access to its most advanced AI models, Claude Fable 5 and Mythos 5, for all users. This decision follows a directive from the U.S. government ordering the company to cut off foreign nationals from using these models, both within and outside the United States, due to national security concerns.
The AI firm stated it received the order at 5:21 p.m. ET, requiring it to block all foreign nationals from accessing the models. Anthropic believes there has been a “misunderstanding” and is actively working to restore access as quickly as possible. The export control order does not impact access to the company’s other AI models.
“We understand the government believes it has identified a way to bypass, or ‘jailbreak’ Fable 5,” the company explained.
“We examined a demonstration of this specific technique being used to uncover a small number of previously known, minor vulnerabilities. These flaws all appear relatively straightforward, and we’ve confirmed that other publicly available models can also identify them without needing a bypass method.”
This sudden action comes just days after the launch of Claude Fable 5 and its counterpart Mythos 5, which runs on the same core model but with certain safeguards relaxed in areas like cybersecurity. The latter, touted as having the “most powerful cybersecurity capabilities of any model in the world,” remains available to a carefully screened group of cybersecurity professionals and critical infrastructure operators.
Anthropic stressed that it has put in place “robust” safeguards to prevent its models from being misused for cybersecurity-related purposes. These protections are supported by a set of safety classifiers designed to detect potential abuse, including jailbreak attempts, and prevent the main model from generating harmful responses.
The cybersecurity classifier is built to block dangerous single-turn requests related to planning cyberattacks, developing exploits, or evading defenses. The company noted that Mythos-class models are highly skilled at discovering and exploiting software vulnerabilities, which could give attackers a significant edge.
Last week, Anthropic disclosed that its Mythos-class model can transform newly reported software vulnerabilities into functional exploits within hours — or even minutes in some cases — rather than weeks, effectively turning N-day vulnerabilities into N-hour threats. The findings indicate that cutting-edge AI models may be equally adept at rapidly weaponizing publicly disclosed flaws.
“A single individual can now convert a month’s worth of security patches into working exploits in just one afternoon — for a few thousand dollars and without any specialized knowledge,” Anthropic’s Red Team stated. “This means the standard patching approach used by software developers today — with monthly release schedules, multi-week staged rollouts, and delays between pre-release and stable versions — is no longer viable.”
Due to Fable 5’s built-in protections, queries on cybersecurity topics will instead be handled by Claude Opus 4.8, the company’s next most capable model.
In its latest statement, the company maintained that no universal jailbreak methods have been developed against its newest models so far, adding that both third-party and internal red-teaming tests have shown its safeguards to be “significantly more effective than those of any previously released model.”
Furthermore, Anthropic argued that “perfect jailbreak resistance” is unattainable for any model provider, since every safeguard used by the industry is vulnerable to non-universal jailbreaks that “work only in very narrow contexts or require extra effort to adapt to each new scenario.”
“So far, the government has only provided us with verbal evidence of a potential narrow, non-universal jailbreak, which essentially involves asking the model to review a specific codebase and fix any software flaws,” the company said.
“We understand that one potential jailbreak was shared with the government. We have reviewed what we believe to be the report underlying the government’s directive and confirmed that the level of capability shown there is widely available from other models (including OpenAI’s GPT-5.5) and is routinely used by the defenders who protect systems every day.”
Anthropic also noted that while it supports the government’s right to block unsafe AI deployments, it argued that the discovery of a “narrow potential jailbreak” should not be grounds for pulling a widely deployed commercial model. The legal process should be “transparent, fair, clear, and based on technical facts,” it added.
Earlier this year, the U.S. Department of Defense designated Anthropic as a “supply chain risk” after the Claude developer attempted to set boundaries around the military use of its technology. The company has filed two lawsuits challenging the designation.
