Technical experts considered the ultimatum of politicians to be a fantasy detached from reality.
The U.S. White House has put forward an ultimatum: the Claude Fable 5 will return to the public only after the company proves that protection from jailbreaks is unbreakable. Security experts are unanimous – it is impossible to achieve this in principle.
The Trump administration has previously recalled Fable 5 from the market, citing the findings of the National Security Agency: the model found ways to circumvent the built-in restrictions in the areas of cybersecurity, chemistry and biology.
Jailbreak is a manipulation of queries that forces the model to ignore its own safety rules. It is the fear that such a tool will be in the hands of unfriendly states that has become an official justification for disconnection.
Anthropic objected: the threat is exaggerated, the consequences of such attacks are minimal. Technical negotiations with the Ministry of Commerce and the Office of the National Cyber Director did not lead to anything. The administration closed the discussion: it no longer intends to argue about the scale of risk.
Now the White House shifts the responsibility to the company as a whole. Officials say bluntly: government agencies do not have the resources to independently look for vulnerabilities in each new model. Anthropic is required to regularly test its developments and the first to inform the authorities of the findings.
The problem is that independent experts consider these requirements technically impossible. According to them, the protective mechanisms of language models are a temporary measure: sooner or later, quite experienced users or more advanced AI systems will find a workaround. There is no fundamental solution – which means that Fable 5 risks remaining inaccessible indefinitely.
While Anthropic and the White House are conducting this dispute, the Ministry of Defense is already actively using AI in combat operations, which previously Anthropic has initially refused. Thus, the director of digital technologies of the department, Cameron Stanley, officially stated that during Operation Epic Fury, the company xAI Hon Elona Musk delivered strikes at more than 2000 targets in 96 hours. A number of senators have already demanded by law to prohibit the use of language models without human control in decisions on the use of force.
The U.S. White House has put forward an ultimatum: the Claude Fable 5 will return to the public only after the company proves that protection from jailbreaks is unbreakable. Security experts are unanimous – it is impossible to achieve this in principle.
The Trump administration has previously recalled Fable 5 from the market, citing the findings of the National Security Agency: the model found ways to circumvent the built-in restrictions in the areas of cybersecurity, chemistry and biology.
Jailbreak is a manipulation of queries that forces the model to ignore its own safety rules. It is the fear that such a tool will be in the hands of unfriendly states that has become an official justification for disconnection.
Anthropic objected: the threat is exaggerated, the consequences of such attacks are minimal. Technical negotiations with the Ministry of Commerce and the Office of the National Cyber Director did not lead to anything. The administration closed the discussion: it no longer intends to argue about the scale of risk.
Now the White House shifts the responsibility to the company as a whole. Officials say bluntly: government agencies do not have the resources to independently look for vulnerabilities in each new model. Anthropic is required to regularly test its developments and the first to inform the authorities of the findings.
The problem is that independent experts consider these requirements technically impossible. According to them, the protective mechanisms of language models are a temporary measure: sooner or later, quite experienced users or more advanced AI systems will find a workaround. There is no fundamental solution – which means that Fable 5 risks remaining inaccessible indefinitely.
While Anthropic and the White House are conducting this dispute, the Ministry of Defense is already actively using AI in combat operations, which previously Anthropic has initially refused. Thus, the director of digital technologies of the department, Cameron Stanley, officially stated that during Operation Epic Fury, the company xAI Hon Elona Musk delivered strikes at more than 2000 targets in 96 hours. A number of senators have already demanded by law to prohibit the use of language models without human control in decisions on the use of force.
