---
title: "Forecasting Trajectory-Level Safety Risks in Black-Box Multi-Turn Interactions"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.26820v1"
permalink: "https://feed.astralor.com/items/itm_a477d1eafaff13f44257558b"
published: "2026-07-29T12:14:23.000Z"
score: 81
channel: "ai"
tags: ["LLM Security","Recast","Multi-Turn Interaction","Risk Forecasting"]
---

# Forecasting Trajectory-Level Safety Risks in Black-Box Multi-Turn Interactions

## AI commentary

> This paper hits the pain point of LLM safety—existing safeguards are all post-hoc, while Recast enables preemptive warning through trajectory-level prediction. It's a must-watch for researchers in AI safety, RLHF, and agent deployment. The experiments are broad and solid, especially the average lead time of 2.41 turns makes proactive intervention feasible.

## At a glance

As large language models evolve from standalone assistants to autonomous agents, safety risks shift from pointwise evaluation to trajectory-level prediction. This paper proposes Recast, a framework that retrieves evidence from short-term dialogues and long-term history via a dual-scale trajectory view, models compositional risk evolution, and uses a causal temporal encoder to predict future risk turns. Experiments across 7 risk categories show 88.3% of future safety failures predicted with an average lead time of 2.41 turns and a false alarm rate of 12.3%. This work shifts from reactive detection to proactive prevention, enhancing LLM safety in complex interactions.

## Key points

- Proposes Recast framework, advancing LLM safety from turn-level violation detection to trajectory-level risk prediction
- Retrieves risk evidence via dual-scale trajectory view (short-term dialogue + long-term history)
- Causal temporal encoder learns latent risk evolution patterns and predicts future risk turns
- Experiments across 7 risk categories: 88.3% success, average lead time 2.41 turns, false alarm rate 12.3%

## Full text

As large language models (LLMs) evolve from standalone assistants into autonomous agents, ensuring their safety requires shifting beyond pointwise risk assessment to understand how risks emerge and unfold over long-horizon trajectories. In multi-turn interactions, malicious intent can be decomposed across seemingly harmless turns and gradually reconstructed through interaction trajectories, eventually resulting in safety failures. Existing safeguards remain largely reactive, detecting manifested violations while lacking the ability to predict latent risk evolution and enable preemptive prevention. To address this limitation, we propose Recast, a safety risk forecasting framework that advances LLM safeguarding beyond turn-level violation detection to trajectory-level risk prediction. Recast first retrieves risk-relevant evidence from both short-term dialogue progression and long-term historical context via a dual-scale trajectory view. It then models compositional risk evolution by capturing the current risk configuration and its temporal dynamics. Finally, a causal temporal encoder learns latent risk evolution patterns and predicts the distribution of future risk emergence turns. Extensive experiments across 7 risk categories show that Recast predicts 88.3% of future safety failures with an average lead time of 2.41 turns, while maintaining a false alarm rate of 12.3%, showcasing the effectiveness of trajectory-level forecasting in identifying emerging risks before safety violations occur.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_a477d1eafaff13f44257558b