Recast Forecasts LLM Safety Risks in Multi-Turn Interactions

As large language models transition from simple assistants to autonomous agents, ensuring their safety requires a more proactive approach than merely detecting violations after they occur. Current safeguards are largely reactive, missing the ability to predict how risks might evolve over extended interactions. Malicious intent can be subtly introduced across multiple turns, gradually building up to a safety failure. To address this, researchers propose Recast, a framework that forecasts safety risks at the trajectory level. Recast first gathers relevant evidence from both the immediate dialogue and long-term historical context using a dual-scale view. It then models how compositional risks evolve by capturing the current risk configuration and its temporal dynamics. Finally, a causal temporal encoder learns these latent risk evolution patterns to predict when future risk emergence turns might occur. Extensive experiments across seven risk categories show Recast can predict 88.3% of future safety failures with an average lead time of 2.41 turns, while maintaining a low false alarm rate of 12.3%. This demonstrates its effectiveness in identifying emerging risks preemptively.