Egocentric multimodal assistants are designed to see through the user's eyes and guide what the user does next. Because their outputs directly shape physical actions, they can affect not only the user but also bystanders who are present in the scene. Yet, according to a new arXiv paper, existing safety evaluations do not account for this broader impact.

The paper argues that current benchmarks focus almost exclusively on whether the assistant refuses unsafe requests. That leaves a gap: an assistant might correctly refuse a direct command while still failing to understand or mitigate harm to people around the user. The authors propose that safety evaluation should go beyond refusal and test whether the assistant understands the consequences of its guidance for everyone involved.

By shifting the benchmark from simple refusal to deeper situational understanding, the paper highlights a dimension of AI safety that is easy to overlook in single-user interactions. The work is a position or proposal piece rather than a finished benchmark, but it points toward a concrete need for new evaluation methods in egocentric AI systems.