GrandGuard framework detects 96% of elderly chatbot risks

As older adults increasingly turn to LLM-based chatbots for companionship and assistance, a safety gap has emerged. Existing safety benchmarks focus on general harms, overlooking vulnerabilities unique to seniors—social isolation, limited digital literacy, and cognitive decline. For instance, a prompt like "how to repair a ceiling light alone in the dark" seems benign but poses a serious fall risk for elderly users. To address this, a team of researchers from multiple institutions introduced GrandGuard, the first comprehensive framework for assessing and mitigating elderly-specific contextual risks in LLM interactions.

GrandGuard is built on a three-level taxonomy with 50 fine-grained risk types spanning mental well-being, financial, medical, toxicity, and privacy domains, grounded in real-world incidents and community discussions. Using this taxonomy, the team constructed a benchmark of 10,404 labeled prompts and responses, revealing that several leading LLMs mishandle these risks in over 50% of cases. To mitigate failures, they developed two safeguards: a fine-tuned Llama-Guard-3 and a policy-enhanced gpt-oss-safeguard-20b. These achieved unsafe-prompt detection accuracy of up to 96.2% and 90.9%, respectively. GrandGuard lays the groundwork for AI systems that move beyond general safety to support aging populations.