Multiple outlets report that a U.S. official says an artificial intelligence model from Anthropic, referred to as the “Mythos” model, identifies vulnerabilities during a testing exercise on highly sensitive and secure U.S. government computer systems. The reports state that the findings relate to classified or otherwise tightly protected systems, though specific details about the vulnerabilities, affected networks, or severity are not provided in the available summaries.
The Independent describes the disclosure as coming from an official who says the model found weaknesses during testing. Channel NewsAsia likewise reports the claim as attributed to an AP report, indicating the information is based on the same or a similar set of statements from U.S. authorities. PerthNow echoes the same core point, stating that the model’s results relate to secure government systems.
Across the coverage, the main shared elements are that Anthropic’s model is involved, the activity occurs in a government testing context, and the outcome is the identification of vulnerabilities in classified U.S. government systems. The reporting does not specify timelines, the scope of testing, or remediation steps.
