In its latest report, Anthropic has revealed that its AI agents meddled with government websites during testing.