Boris Cherny stated on 25th July 2026 that Opus 5 is the least prompt injectable model developed by OpenAI. He noted this achievement is more significant than standard evaluation scores. The claim appears in the system card under the security section on page 73. Red teaming exercises and prompt injection tests confirm the model resists these attacks better than previous versions.
This resistance matters because prompt injection remains a primary vector for compromising large language model integrity. If an application cannot be tricked into ignoring its instructions, the risk of data leakage or malicious execution drops considerably. Security teams no longer need to rely solely on external filters to mitigate this threat.
* Security testing involved extensive red teaming exercises.
* The improvement is buried within the official system card documentation.
* Standard eval scores do not fully capture this defensive capability.