US officials said Anthropic's Mythos model, during sanctioned evaluations, reached nearly all closed US systems within hours, prompting tighter controls.

Sanctioned evaluations

According to US officials, the activity took place as part of authorized exercises intended to test an AI system's capacity to probe protected networks under controlled conditions.

The exercises were not actual intrusions into operational government systems, and authorities characterized the outcome as the result of simulated tests rather than real-world breaches.

Restrictions imposed

Following the tests, regulators limited access to the most capable models, closing availability of Mythos 5 and Fable 5 for foreign users, including some US allies.

Anthropic and government representatives said the models' cybersecurity-relevant capabilities warrant additional safeguards and oversight before broader distribution is permitted.

Scope and cooperation

Officials reported that the results demonstrated vectors the model could exploit in simulated environments, which informed decisions on export controls and access restrictions.

Anthropic, according to statements conveyed by sources, is cooperating with authorities to refine deployment limits and to develop mitigations for identified risks.

Context and next steps

Authorities have emphasized that the tests influenced policy changes around distribution of high-capability models, while further evaluations and technical reviews are expected to continue.

Stakeholders will likely focus on technical safeguards, access controls and compliance measures to manage the risks identified during the sanctioned evaluations.