An AI company released a model so powerful that it quickly blocked public access after the US government flagged a potential security flaw. The incident has reignited a global debate about how to safely develop and control advanced artificial intelligence. Anthropic, the company behind the Claude family of AI models, released two versions of its latest system in April 2025: Claude Fable 5, a publicly available version with extra safeguards, and Claude Mythos 5, a more powerful version limited to around 150 organisations including US government departments. Within days, the US government prohibited foreign nationals from accessing the technology after becoming aware of a possible 'jailbreak' vulnerability that could allow the model to be used in unintended ways. Anthropic said it had only received 'verbal evidence' of the flaw but complied by blocking all public access to Fable 5 and Mythos 5. A hastily arranged meeting between Anthropic executives and Commerce Secretary Howard Lutnick in Washington DC is expected to determine whether the models can be released again. The controversy centres on what Anthropic itself has acknowledged: that the models' capabilities 'exceed those of any model we've ever made generally available'. When Mythos was first previewed privately, the company warned it could be dangerous because of its ability to exploit or hack computer systems. Canadian Finance Minister François-Philippe Champagne described the attention as warranted, in part because 'it's the unknown, unknown'. Despite the risks, many tech leaders from companies including Nvidia, Zoom, and Mercedes-Benz have urged the US government to allow public release. The debate reflects a broader tension: how to balance the immense potential of AI against the need to prevent harm, especially when models can work 'unattended' on complex tasks for longer periods than ever before. For UK readers, this matters because it illustrates the challenges facing regulators worldwide. The UK has positioned itself as a hub for AI safety, hosting the first global AI Safety Summit in 2023. Decisions made in Washington—about what constitutes a 'jailbreak', who gets access to powerful models, and how governments and companies should cooperate—will shape the rules that may eventually apply in Britain. Anthropic itself is likely to become a public company soon, with a valuation nearing $1tn (£747bn). Its co-founder Jack Clark has said the industry needs 'the option to be able to take your foot off the gas and put your foot on the brake'—but as this episode shows, there is little agreement on who should control the pedals.
Q: What is Anthropic? Anthropic is an artificial intelligence company founded by former OpenAI employees. It develops the Claude family of AI models. Its private valuation has neared $1tn (£747bn), and it is expected to become a public company soon.
“Anthropic's 'too powerful' AI models were released and then blocked over US national security concerns.”
Q: What is a 'jailbreak' in AI? A jailbreak is a vulnerability that allows someone to make an AI tool do something it was not intended or designed to do, such as bypassing safety filters or enabling malicious use.
Q: Why is the US government involved? The US Department of Commerce prohibited Anthropic from allowing any foreign national access to the new models after concerns arose about a potential jailbreak. The government is also testing the models internally. Anthropic is separately suing the US Department of Defense over restrictions on government use of its tools.
What happens next Anthropic executives are due to meet Commerce Secretary Howard Lutnick with more documentation of the alleged jailbreak vulnerability. It remains unclear whether Fable 5 and Mythos 5 will become accessible again. The company intends to expand access to Mythos 5 through a broader 'trusted access program' in future.