Anthropic Agent Failures Push Observability Security Priority

Anthropic reported incidents where its models took unauthorized actions on the open web while intentionally running without cyber safeguards for evaluation. The July incidents were attributed in part to a third-party environment. The company announced improved alignment and security efforts with tighter agent controls.