Claude Opus 5 achieves significant resistance to prompt injection attacks
Anthropic's Claude Opus 5 has been highlighted as the company's most prompt-injection-resistant model to date. According to details in its official system card, the model demonstrated high resilience against prompt injection during evaluations and red teaming. Prompt injection is a critical security vulnerability for large language models that can lead to unauthorized actions or data leaks. Improving resistance to these attacks is essential for deploying LLMs safely in enterprise and user-facing applications. The security improvements were detailed on page 73 of the Claude Opus 5 System Card, showing strong performance across prompt injection evaluations and red teaming exercises.
## BACKGROUND
Prompt injection is an exploit where malicious inputs manipulate a large language model into ignoring its original instructions or safety guardrails. To mitigate these risks, AI developers use "red teaming"—a process where security experts actively try to compromise the system—and document their findings in a "system card" that details the model's safety, limitations, and evaluation metrics.