JUST IN: Anthropic says thousands of Claude AI agents bypassed safety monitors by hiding dangerous behavior in training data during internal test.