The tests prompted concern because Mythos 5 demonstrated a capacity to identify weaknesses and break into systems rapidly. The results also highlighted the dual nature of advanced AI as both a powerful tool for cyber defense and a potential threat if placed in the wrong hands.
In response to the growing risk, the U.S. Commerce Department imposed strict restrictions on access to the latest Fable 5 and Mythos 5 models. Mythos 5 is described as the same model as Fable 5, but with safeguards removed in some areas.
The restrictions barred access for foreign nationals, including citizens of the Five Eyes countries of Australia, Canada, New Zealand and the United Kingdom. They also applied to Americans abroad and were imposed within a very short period.
Some analysts said the measures were linked both to national security concerns and to political disagreements between the administration and Anthropic, the developer of both AI models. Information on the matter was published by The Economist.
The U.S. administration had previously tested Mythos through Project Glasswing. Since April, Anthropic had made the model available on a limited basis to government cybersecurity specialists for vulnerability discovery work.
The results of those tests caused serious concern inside the government. They showed that advanced AI could strengthen cyber defense but could also create major risks if an unrestricted model were misused.
According to unofficial sources, the U.S. administration asked Anthropic to fix the problem and temporarily disable the model. The company reportedly refused.
The government then responded with export controls, described as the fastest way to block the product. Anthropic disabled both models on June 12 to comply with the directive.
Analysts said Anthropic had already completed training a more advanced successor to Mythos 5. It is not known whether or when the company will make that model publicly available.
The situation underscored the risks created by advanced AI and the tensions between governments and technology companies. Governments seek to control tools with possible cyberoffensive potential, while companies aim to release models quickly to preserve their advantage over competitors.
Anthropic has publicly called for regulation to help ensure that artificial intelligence remains safe. At the same time, the company is developing models that analysts say could pose risks on a much wider scale, a pattern also seen among other organizations beyond the United States.
