Quoting Boris Cherny

Summary

Boris Cherny notes that Opus 5 is the least susceptible to prompt injection model to date, based on its performance demonstrated through PI evaluations and red team exercises.

Why it matters

Prompt injection attacks can steer language models to produce unwanted or harmful outputs; therefore, a model resistant to such injections is crucial for reliable AI deployment. Opus 5's robustness signals a step forward in protecting LLM interactions from malicious prompts.

Highlights

• Opus 5 demonstrates strong resistance to prompt injection across multiple evaluations.
• This claim is documented in the model's system card (page 73).
• Red team exercises confirm that successfully injecting prompts into the model is difficult.

Critical take

The evaluation may be based on specific test scenarios; broader real-world usage could reveal injection vectors not covered in the current assessments.