AnthropicSynthID‑Text watermarking can alter Claude safety behavior under adversarial prompts3 min read·1mo ago·via Ars Technica Technology Lab