A first-person editorial on a sudden Claude quality breakdown: over-refusal of defensive cyber work, fabricated threat-intel output, repeated hallucination after apology, and the broader risk that safety overreach can damage product trust.