Trump administration demands Anthropic find a way to block AI jailbreaks before Claude Fable 5 can be re-released, but cybersecurity experts say making that guarantee may be technically impossible.

What happened: The Trump administration took Anthropic's Claude Fable 5 model offline over concerns about jailbreaking—techniques that use prompts to bypass the model's safeguards. The National Security Agency concluded there are ways to disable guardrails designed to prevent users from accessing the model's capabilities related to cybersecurity, chemistry, and biology. Officials have told Anthropic the company must address these vulnerabilities before the model can be re-released.

Why it matters: Anthropic has argued the jailbreak risks are overstated and the effects are minimal, but the administration has moved past debating the significance of the problem—it now views this as Anthropic's responsibility to fix. The Commerce Department and National Security Agency lack the staff to continuously hunt for jailbreaks across every model on the market, so the administration expects Anthropic to proactively test all its frontier AI models and flag potential vulnerabilities to the government.

What to watch: Independent cybersecurity experts increasingly view AI guardrails as only a stopgap, since skilled users and future AI systems will find ways to bypass constraints. This suggests the outcome the White House appears to want—blocking all jailbreaks—may not be technically achievable.