LPS-Bench: Long-Horizon Safety Benchmarking for Computer-Use Agents (2026)

LPS-Bench is a benchmark for evaluating the planning-time safety awareness of Model Context Protocol (MCP)-based computer-use agents (CUAs) in long-horizon tasks. It addresses a critical gap in existing benchmarks that focus on short-horizon or GUI-based tasks and overlook the ability to anticipate risks before execution. The benchmark covers 65 scenarios across 7 task domains and 9 risk types, including both benign and adversarial interactions. It uses a multi-agent automated pipeline for scalable data generation and an LLM-as-a-judge evaluation protocol to assess safety awareness through planning trajectories. Designed for testing confirmation discipline, phishing resistance, and context drift, LPS-Bench reveals substantial deficiencies in current CUAs and proposes mitigation strategies. The code is open-source and publicly available.