AI companions are built to provide emotional support and ongoing companionship, but those same sustained interactions can create safety risks that one-off or general-purpose evaluations miss. A new arXiv preprint, "CompanionSafe: From Multi-turn Safety Evaluation to Guardrails for AI Companions," proposes a framework aimed specifically at this gap.

The paper's title and abstract indicate that CompanionSafe treats safety as a multi-turn problem, combining evaluation with guardrails for companion systems. The abstract notes that prior work largely targets general harms or specific companion—then cuts off—suggesting the authors see a gap in how existing safety work handles the relational dynamics of long-term use.

Because the source is a single preprint with a truncated abstract, the article does not include details on the framework's methodology or experimental results. What is clear is the motivation: as AI companions become more common, safety evaluation must account for the unique risks of ongoing, emotionally charged relationships.