Quoting Boris Cherny
Anthropic's Opus 5 is their most prompt-injection-resistant model to date.
“Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully.”
Boris Cherny, likely an Anthropic engineer, highlights that Claude Opus 5 achieves a new bar for prompt injection resistance, backed by PI evals and red teaming documented in the model's system card. This matters because prompt injection is one of the primary attack vectors against LLM-powered agents and agentic workflows, making robustness here a significant trust and safety milestone. It signals Anthropic is treating security properties as a first-class differentiator alongside capability benchmarks.