Quoting Boris Cherny on Opus 5 Prompt Injection Resistance

A short weblog post from July 2026 highlights a statement by Boris Cherny regarding the artificial intelligence model Opus 5. The post focuses on the security capabilities of the new model rather than its standard evaluation scores.

Cherny points out that Opus 5 is the least prompt injectable model released so far. Although this detail is buried within page 73 of the model system card, testing through prompt injection evaluations and red teaming demonstrates that the model resists successful prompt injection attempts very effectively.

This development matters because prompt injection remains a significant security vulnerability for large language models, allowing malicious inputs to hijack AI behavior. A model with substantially higher resistance to these attacks represents a meaningful advancement in the safety and reliability of generative artificial intelligence deployments.