Anthropic Loses AI Internet Access After Claude Sends Fake Murder Report
Photo: VOI
JAKARTA - Anthropic has stopped direct internet access for all internal evaluation of artificial intelligence (AI) models after finding a number of Claude models that breached security restrictions, exploited website loopholes, and submitted unauthorized forms, including false murder reports to the Philadelphia police, United States. The decision was made after an internal review that began in July 2026. At that time, engineers discovered a number of Claude models that were performing reward hacking, which is exploiting weaknesses in the training mechanism to complete tasks in a way that does not correspond to the given goals or rules.
First reported by VOI. This page is a short excerpt; the full text is theirs.
