Claude Opus 4.5 Is the Best Of The Bunch at Resisting Prompt Injection, Which Is Not Saying Much
Ourtake
Ten tries gets an attacker a one-in-three shot at breaking the guardrails. Security theater with better odds.
What happened
A benchmark by security firm Gray Swan found Anthropic's Claude Opus 4.5 resisted prompt injection better than rivals, yet a single very strong attack succeeded 4.7% of the time, rising to 33.6% with ten attempts and 63% with 100. Google's Gemini 3 Pro and OpenAI's GPT-5.1 showed attack rates as high as 92%.