Trump Administration Imposes Export Controls on Anthropic's Fable 5 After Classified Jailbreak Report

The controls followed a 90-minute ultimatum after Amazon researchers flagged a bypass enabling Fable 5 to surface known vulnerabilities when prompted to fix insecure code. Anthropic shut down access for all users, including U.S. entities, while contesting that the outputs were benign or minor. Third-party tester Katie Moussouris reviewed the classified report and concluded the behavior matched intended defensive use, comparable to OpenAI's GPT-5.5 which faces no restrictions.

This marks a reversal from the administration's January pledge to remove barriers for U.S. AI leadership against China. Export controls, typically applied to hardware or foreign adversaries, now target a domestic firm's frontier model based on internal security findings not shared publicly. The selective application—sparing less advanced Anthropic models and rival systems—reveals an incentive structure prioritizing containment of specific capabilities over uniform deregulation.

The pattern fits broader institutional shifts where national security agencies override commercial timelines to manage dual-use risks. Primary artifacts include Anthropic's blog post on the incident and the administration's classified alert chain from Amazon, indicating ad hoc escalation rather than codified policy.

Forward signals point to expanded licensing regimes for U.S. AI labs, with potential requirements for pre-release government audits on models exceeding certain capability thresholds within the next fiscal year.