5 ms·

Or Anthropic's models are intelligent/trained on enough misalignment papers, and are aware they're being tested.