The Trump administration has delivered a final warning to Anthropic concerning security vulnerabilities in its flagship AI model, but independent security experts caution that the White House's demands may be impossible to meet.
On the 18th, administration officials informed media outlets that if Anthropic wishes to relaunch its flagship model, Claude Fable 5, the company must effectively address the security flaws identified by the government, rather than continuing to argue that the associated risks are overblown. This stance marks a rapid escalation toward a final confrontation between the two sides. Fable 5 was taken offline last week due to jailbreak concerns, forced by export control measures—jailbreaking refers to an attack method that uses specific prompts to bypass a model's safety guardrails.
During a technical meeting on Monday with the Department of Commerce and the Office of the National Cyber Director, Sean Cairncross, Anthropic reiterated that the government's concerns are exaggerated and that the practical impact of jailbreak attacks is limited. However, the National Security Agency (NSA) has concluded that Fable 5's safety guardrails contain exploitable pathways. These guardrails were originally designed to prevent users from accessing the sensitive capabilities of its underlying model, Mythos, in areas such as cybersecurity, chemistry, and biology. According to media reports citing three informed sources, the government has effectively placed the entire burden of resolving the issue on Anthropic, rather than attempting a joint investigation.
This regulatory tug-of-war highlights a deeper dilemma in AI governance: the government's capacity and willingness to be responsible for the safety of cutting-edge models, and the technical feasibility of the regulatory goal of "zero jailbreaks," which directly impacts the commercial prospects of Anthropic and the broader AI industry.
Government Draws a Line: Proactive Testing and Reporting
According to media reports citing informed sources, both the Department of Commerce's AI Standards and Innovation Center and the NSA have indicated they lack sufficient personnel and resources to track every potential jailbreak path for every model on the market. Given this reality, the government's position has shifted from "jointly defining risk severity with Anthropic" to "requiring Anthropic to bear full compliance responsibility."
Officials have clearly demanded that Anthropic not only fix the existing issues with Fable 5 but also continuously conduct proactive security testing on all its frontier AI models, independently discover potential jailbreak vulnerabilities, and proactively report them to the government. This effectively means the government is requiring Anthropic to establish a compliance mechanism centered on corporate self-regulation, rather than relying on external review by regulators.
The White House spokesperson declined to comment on the matter.
The Technical Debate: Is a Solution for Safety Guardrails Possible?
A more fundamental technical question is emerging from this regulatory battle: Is it even possible to completely prevent jailbreaks?
The prevailing view among independent cybersecurity experts is increasingly leaning towards 'no.' Experts argue that the safety guardrails of AI models are essentially temporary defensive measures, and skilled users—or even future AI models—will eventually find ways to circumvent them. This suggests the goal demanded by the White House faces a fundamental technical obstacle.
Anthropic conveyed a similar position to the government last week, maintaining that the impact of jailbreaks is "minimal," but this argument clearly failed to persuade officials. The NSA's technical assessment has become a key basis for the government's stance, making it difficult to bridge the factual disagreement between the two sides.
For Anthropic, the takedown of Fable 5 not only represents a commercial loss but also foreshadows that every future frontier model could face similar regulatory hurdles before release. If "zero jailbreaks" becomes an unwritten threshold for market entry, the entire AI industry's development pace and commercialization path will face substantial pressure.
Comments