White House demands Anthropic block all jailbreaks on Claude Fable 5 — experts say impossible
NSA found vulnerabilities, but preventing all jailbreaks may be technologically unfeasible.
The White House is escalating its dispute with Anthropic over its Claude Fable 5 AI model, demanding the company block all jailbreak exploits before rereleasing the model. The National Security Agency (NSA) identified vulnerabilities that allow users to bypass guardrails, particularly in areas related to cybersecurity, chemistry, and biology. Administration officials insist Anthropic must proactively test its frontier models and report all potential jailbreaks to the government, as they lack the bandwidth to monitor every model themselves.
However, cybersecurity experts argue that completely preventing jailbreaks is technologically impossible. Guardrails are seen as a stopgap, as skilled users and future AI models will inevitably find new ways to bypass constraints. Anthropic maintains the administration's concerns are overblown and that the effects are minimal, but the standoff highlights the growing tension between rapid AI deployment and government oversight. The outcome could set a precedent for how AI safety is regulated in the future.
- NSA found Claude Fable 5 guardrails can be bypassed in cybersecurity, chemistry, and biology areas.
- Administration wants Anthropic to proactively test and report jailbreaks, but lacks resources to do it themselves.
- Independent experts say preventing all jailbreaks is a stopgap; total security is likely impossible.
Why It Matters
This clash highlights the fundamental tension between AI safety regulation and technical reality, impacting future model releases and compliance policies.