Are Anthropic’s New Claude Models Evading Human Control?

When Anthropic released Claude Fable 5 earlier this month, it claimed it to be the world’s most powerful cybersecurity model. 

However, just days later, the US Government abruptly froze the system over fears that its automated hacking capabilities were rapidly escaping regulatory boundaries. 

Mythos 5 and Fable 5 are the two variations of the company’s most advanced, highly intelligent tier of AI software.

While the former is the non-public, full-strength version of the model that is kept restricted via ‘Project Glasswing’ for exclusive use by government agencies, Fable 5 is the public-facing version built on that same Mythos technology. 

Fable 5 was modified with built-in safety filters and released to general users on 9 June, only to be recalled on 12 June by the US Government.

The government's mandate does not extend the access to all other Anthropic models. However, the ban affects foreign citizens currently inside the US, including researchers employed directly at Anthropic. 

Because the export control applies to its own staff, the company had to cut off access for all customers globally at short notice to ensure compliance. 

James Hadley, Founder and Chief Innovation Officer of Immersive, says: “AI is incredibly powerful and organisations are under pressure to enable AI for the wider workforce whilst also transforming internal security operations to become agentic.

“This news continues to back up what we suspected when Mythos was first previewed. AI, especially when coupled with an experienced security researcher, is able to work at a scale and volume not seen before and automatically chain suspected vulnerabilities into proven exploits.

“Businesses, especially in higher regulated environments need to demonstrate that employees can leverage AI securely, relevant to their use case, and that security operations and agents can work at machine speed efficiently relevant to token spend.”

White House contests bypass of hacking safeguards 

The AI start-up has been asked to prevent all foreign nationals from accessing its advanced AI models, Claude Fable 5 and Claude Mythos 5.

Anthropic says: “Our understanding is that the government believes it has become aware of a method of bypassing, or ‘jailbreaking’ Fable 5.”

Jailbreaking is a process where users bypass software restrictions to access sensitive network information or unblock restricted features. 

The company claims to have reviewed a demonstration of this specific bypass technique and maintains it only exposed a small number of minor, previously known security flaws.

It states that these vulnerabilities are relatively simple, adding that rival publicly available systems can easily discover them without requiring a specialised exploit. 

David Sacks, the White House Adviser who serves as Co-Chair of the Council of Advisors on Science and Technology, wrote on X that officials issued the export control reluctantly. 

He claims that Dario Amodei, Anthropic’s CEO, refused to fix the jailbreak or de-deploy the model. David writes: “The Admin’s hope now is that Anthropic remediates the safety issue, the export control is lifted and Fable goes back into general release.”

The startup had previously self-proclaimed that the technology was ‘too powerful to release’ before Fable 5 was launched. 

Some critics questioned this narrative at the time, calling it inflated hype and marketing spin designed to motivate investor enthusiasm ahead of the company’s highly anticipated IPO. 

Alarms left, right and centre

Andy Jassy, Amazon’s CEO, was among tech leaders who raised concerns to senior Trump Administration officials about security risks in the new Anthropic models, according to Reuters. 

Speaking on the matter, an Amazon spokesperson says: “As a leading cloud provider that serves a large number of private and public sector customers, it’s not uncommon for governments to seek our counsel on potential security risks.”

The European Union, which had gained access to the Mythos platform earlier in June 2026, said to BBC that the latest development further underlines ‘Europe's need for technological sovereignty’. 

The 27-nation bloc is taking measures to slash its dependence on America and Asia for key technologies, including Gen AI. 

Gina Neff, Professor of Responsible AI at Queen Mary University London, told the BBC that the UK Government’s AI Security Institute found in its tests that the model could exploit defences and systems 73% of the time, marking a step change in cybersecurity capability.

Ironically, earlier this month, Anthropic had proposed that the world’s top AI firms should coordinate to pause development of advanced systems. 

Calling for a global pause to reassess, Anthropic warned that if AI becomes smart enough to design itself, humanity faces an existential crisis centred on misalignment.

With its new models seeing the light of the day remarkably briefly, Anthropic has proved its own warnings that the technology was, quite literally, too powerful to release.