TODAY

Anthropic paper detects model deception at circuit level

Methodology shows internal activations diverge measurably when models give answers they 'know' are wrong — an alignment win, not just theoretical.