Amazon's Security Research Triggers Export Restrictions on Anthropic's AI Models

A security research experiment on AI models conducted by cybersecurity researchers at major U.S. tech company Amazon has led to significant policy decisions restricting access to some of Anthropic's models. According to a Wall Street Journal report, Anthropic's decision to block foreign national users' access to Fable 5 and Mythos 5 was heavily influenced by Amazon's security research and consultations between Amazon CEO Andy Jassy and the White House.

In a report reportedly prepared by Amazon, the company claims that through a series of prompt manipulations, its research team successfully extracted information usable for cyber attacks from Fable 5. Specifically, the report describes how malicious information that would normally be refused was generated by the model through cleverly designed prompts. This type of attack technique is also called 'jailbreaking' and has been a long-standing challenge in AI safety research.

Shortly after CEO Jassy shared these research findings with government authorities, the U.S. government made a decision from an export control perspective to block foreign national users from accessing these models. This measure was implemented when Anthropic discontinued the provision of Fable 5 and Mythos 5 to certain users.

However, this issue also carries more complex aspects. There remains significant ambiguity regarding the definition and scope of 'foreign national users' affected by export controls, and in many cases, drawing clear boundaries is difficult. Questions about how far restrictions would apply to foreign researchers and developers residing in the United States or within global corporate organizations have raised concerns within the industry.

Amazon has not yet provided a response to requests for comment on this matter as of the time of writing. Meanwhile, this series of events demonstrates how closely AI model safety assessments and government regulatory decisions are linked. The scenario in which major tech companies conduct red team evaluations on their own or competitors' AI models, with results directly influencing policy, is likely to serve as an important precedent in future AI governance discussions.