Anthropic Discloses Opus 5 Safety Depends on System Prompts

Anthropic's Claude Opus 5 system card, published July 24, 2026, reveals the model's harmless response rate on the API dips below recent baselines when no system prompt is provided. That safety behavior depends on surrounding scaffolding, not just the model itself. API users overriding default prompts should weigh this gap in their risk assessments.