White House Demands Anthropic Block All Jailbreaks as Impasse Over Claude Fable 5 Intensifies

Summary

The Trump administration has escalated pressure on Anthropic following the export control suspension of Claude Fable 5, the company's most advanced model, over concerns that hackers could use prompt-based jailbreaks to circumvent the model's safety guardrails. Trump officials have made clear that Anthropic cannot rerelease the model without demonstrably addressing what the National Security Agency has confirmed are exploitable vulnerabilities, particularly those restricting access to dangerous capabilities in cybersecurity, chemistry, and biology.

While Anthropic has downplayed the severity of jailbreaking risks, the administration has moved beyond debating significance and now views the problem as Anthropic's responsibility to solve. Officials are pushing the company to adopt more proactive security testing for all frontier models and to self-report vulnerabilities before public release. However, the feasibility of what the White House is asking remains deeply contested among independent cybersecurity experts.

Security researchers increasingly argue that AI guardrails are merely a stopgap solution, suggesting that skilled users and advanced AI systems will inevitably find ways to bypass any safeguards. This raises a fundamental question: whether completely blocking jailbreaks is a realistic goal or simply a risk that must be managed alongside other security measures.

• Cybersecurity experts question whether preventing jailbreaks entirely is even technically possible, suggesting guardrails are inherently bypassable
• The dispute reflects growing tension between government AI safety demands and industry views on the technical feasibility and scope of those demands