Addendum to GPT-5 System Card: Sensitive conversations

OpenAI has published an addendum to the GPT-5 system card detailing improvements in handling sensitive conversations. The update introduces benchmarks for emotional reliance, mental health interactions, and jailbreak resistance. According to OpenAI Blog, these benchmarks measure how well the model navigates delicate topics without causing harm or being manipulated. For developers building AI workflows, this means more predictable guardrails when deploying GPT-5 in applications that involve mental health support, emotional advice, or other sensitive domains. The jailbreak resistance benchmarks also indicate reduced vulnerability to prompt injection attacks. Practically, builders can use these benchmarks to validate safety in their own systems and adjust prompt engineering or oversight accordingly. The update reflects ongoing efforts to balance capability with safety in large language models.