Hacker News discussion: Anthropic's Opus 5 Is Better at Resisting Prompt Injection

Anthropic’s Opus 5 reportedly reduced attacker success on the IPI prompt-injection benchmark to 2.0% within 15 attempts, down from 5.5% for Opus 4.8, and to 0.2% in a single attempt. It was the most robust model evaluated, outperforming other Claude and non-Claude models; the strongest non-Claude model had a 16.5% success rate within 15 attempts, while GPT 5.6 variants ranged from 20.0% to 43.9%. The results suggest improved resistance in tested scenarios, though the article notes that preventing prompt injection is impossible in the general case.