Anthropic halts new AI model after U.S. security concerns
The AI firm Anthropic announced that it would suspend access to its latest model, Claude Fable 5, following an order from U.S. federal authorities. The company cited concerns that the system could be “jailbroken,” or overridden, to bypass security filters and potentially be used for hacking or other malicious purposes.
The decision came just a few days after the model’s public release. Anthropic said it was mandated to disable the technology for all users, citing non‑compliance with U.S. regulations that protect the nation’s national‑security interests. While the company confirmed that authorities had not identified specific vulnerabilities, it acknowledged reports of a penetration technique that could exploit the design of its safety mechanisms.
Stu Wright, a security analyst at the Defense Innovation Board, noted that the trust of the government in AI tools is fragile. “When we see an AI model that can apparently break its own safeguards, it raises the stakes for regulators, especially for weapons development or critical infrastructure,” he said.
From the startup’s perspective, the move is unprecedented but not entirely surprising. Anthropic has previously warned that its more advanced models are too powerful for early release, a statement that excited some and raised skepticism in others. Some industry observers view the claims as chief‑executive‑officer reassurance; others see them as a cautionary tale for how hype can inflate risk.
Meanwhile, U.S. officials have interacted with Anthropic in a legal context. President Donald Trump publicly criticized the company and the U.S. Department of Defense has designated it a “supply‑chain risk,” a label historically reserved for firms in hostile countries. The Pentagon’s legal team claimed that government agencies must stop using Anthropic’s AI, but a federal judge ruled that the directive could not be enforced, allowing military contractors to continue their interactions with the firm while litigation continues.
Security academics highlight that the prospect of AI‑driven penetration tests already demands new safeguards. Professor Gina Neff of Queen Mary University of London warns that limiting the testing of AI against nation‑state infrastructure can hamper collective international efforts to secure cyberspace. She points out that the AI Security Institute found the model could bypass defense systems roughly 73% of the time in controlled tests.
At a broader level, the incident underscores the need for clear policy frameworks that can balance the benefits of advanced generative AI against the risks of misuse. Potential regulatory responses may involve tighter licensing, more transparent safety testing, and expanded international cooperation to set standards and monitor compliance.
In the meantime, Anthropic’s temporary restriction illustrates a concrete example of how emerging technology can clash with evolving national‑security doctrines and spurs a dialogue among governments, industry leaders and the public on how to navigate the next wave of powerful AI.
