WorldRiskBench: Evaluating LLM Safety Against Real-World Triggered Risks

WorldRiskBench is a streaming benchmark designed to evaluate the safety of large language models and AI agents. It models safety as a world-dependent property, where risk emerges from the interaction between user instructions and real-world events. The dataset was authored by SsssOvO and last updated on May 18, 2026.

Use Cases

• Benchmarking LLM safety based on event-driven risk scenarios described in the overview.
• Evaluating AI agent safety based on the world-dependent property of risk.
• Comparing model performance against conventional static safety benchmarks based on the described streaming methodology.

Strengths

• Focuses on real-world, event-driven risk scenarios, a distinct approach from conventional benchmarks.
• Models safety as a dynamic, world-dependent property rather than a static prompt property.
• Designed as a streaming benchmark, suggesting potential for continuous updates.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Freshness

Last updated 2026-05-18 08:22:36; freshness should be verified.