Deep Tide TechFlow message: On July 25, Anthropic said in its Opus 5 system card that the model is the least susceptible to prompt-injection attacks of any it has released to date. Based on the prompt-injection evaluations and red-team test results, Opus 5 demonstrates stronger resistance against malicious prompt injections. Prompt injection is one of the core risks in AI security: attackers use carefully designed inputs to bypass a model’s safety constraints and induce it to perform unintended behavior. Anthropic published the relevant evaluation details on page 73 of its system card.