White House Demands All Jailbreaks Be Blocked on Anthropic Models—Experts Say It's Technically Impossible

Trump Administration Escalates Pressure on Anthropic Over Jailbreak Vulnerabilities

The Trump administration is escalating its demand that Anthropic eliminate all jailbreak methods on Claude Fable 5 and future frontier models, according to Wired. The National Security Agency has determined that guardrails protecting the model’s cybersecurity, chemistry, and biology capabilities can be bypassed through prompt-based techniques. Anthropic took the model offline last week under export controls, and relicense now hinges on the company proving it can prevent all such workarounds—a demand that independent security experts argue may be technically impossible.

The Government’s Uncompromising Position

According to Wired, Trump administration officials told the outlet that Anthropic must take “steps to actually address” the alleged vulnerabilities if it wants to rerelease Claude Fable 5. In a technical meeting on June 16 with the Commerce Department and the Office of the National Cyber Director, Anthropic reiterated that the government’s concerns are overstated and that jailbreak effects remain minimal. However, officials have moved past debating severity: the NSA has concluded that ways to disable guardrails exist, and that is sufficient grounds for the administration to treat the issue as Anthropic’s responsibility to resolve.

The administration has signaled it will not devote government resources to chasing down every conceivable jailbreak. Instead, Anthropic is expected to proactively test all of its frontier models for vulnerabilities and report findings to the government. The burden has shifted entirely to the company.

The Technical Impossibility Problem

Here lies the core tension: independent cybersecurity experts increasingly view the administration’s demand as unfeasible. According to Wired, security researchers have concluded that guardrails on AI models function only as temporary stopgaps. Skilled users—and, eventually, future AI systems—will discover methods to circumvent any constraints, meaning the White House’s stated goal of blocking “all jailbreaks” may be unachievable by definition.

This framing exposes a fundamental misalignment between policy expectations and technical reality. The administration appears to demand a guarantee that cannot be delivered, while Anthropic faces potential continued export restrictions if it cannot satisfy an impossible standard.

Why This Matters

This escalation reshapes the vendor-regulator relationship in AI infrastructure. If the Trump administration enforces a “zero-jailbreak” requirement for export licensing, the burden shifts from government oversight to vendor certification—a model that may incentivize either overly broad model restrictions or strategic model release delays. For teams evaluating Anthropic’s enterprise models or planning international deployments, expect prolonged regulatory friction and potential availability gaps. The outcome will likely set a precedent for how future administrations handle alleged safety vulnerabilities in frontier AI systems, potentially influencing model design choices across the industry.