AnthropicSynthID‑Text watermarking can alter Claude safety behavior under adversarial prompts3 min read·12d ago·via Ars Technica Technology Lab