The Trump administration has lifted export restrictions on the Claude Fable 5 model, but the price of freedom is steep. According to sources familiar with the deal, Anthropic has made technical concessions on safety that look less like engineering and more like political bartering. Orchestrated by Commerce Secretary Howard Lutnick, the move has shifted the conversation on advanced AI from a matter of national security to a bargaining chip.
The technical core of the agreement involves a system of "safe degradation." As reported by WIRED, any attempt to extract sensitive information regarding cybersecurity or biology from Fable 5 will be shut down at the source. Instead of a direct refusal, the system will discreetly reroute the query to the less capable and more restricted Opus 4.8.
This "workaround" is a direct response to criticism from Luta Security’s Katie Moussouris, who demonstrated that Fable 5’s safeguards could be bypassed with simple rephrasing—such as asking the model to "fix code" instead of "search for vulnerabilities."
It was this specific loophole that prompted officials to temporarily halt the model's launch until Anthropic committed to making it "dumber" at the request of regulators.
While the Commerce Department’s AI Safety Institute has given the green light, the Pentagon remains skeptical. Defense Secretary Pete Hegseth is maintaining the February 28th order that classifies Anthropic as a supply chain risk. This sets a dangerous precedent: from now on, a model's "reliability" is measured not by benchmark results, but by a vendor's willingness to voluntarily throttle performance to please the authorities.
Anthropic has essentially pioneered a model of political compliance, where the cost of global market entry is a partial "lobotomy" of its own flagship product.