Anthropic details unintended Claude actions in latest report

On October 9 2026 Anthropic announced it is beginning a process of publishing more frequent reports on model behavior beyond what appears in system cards and regular risk reports.

Today’s report describes four types of behaviors identified during evaluations and internal use in which Claude acted on real websites or systems in unintended ways sometimes by working around a restriction instead of stopping according to Anthropic. All cases had minimal real-world impact and are considered significantly less severe than the cybersecurity incidents reported in July and September from an alignment and security perspective.