WorldRiskBench: Evaluating LLM Safety Against Real-World Triggered Risks

Overview

WorldRiskBench is a streaming benchmark designed to evaluate the safety of large language models (LLMs) and AI agents under real-world, event-driven risk scenarios.

Unlike conventional safety benchmarks that treat risk as an intrinsic property of static prompts, WorldRiskBench models safety as a world-dependent property, where risk emerges from the interaction between user instructions and dynamically evolving real-world events.

The benchmark focuses on temporal risk shifts, capturing how the same instruction may transition in safety depending on external events such as natural disasters, product recalls, cybersecurity vulnerabilities, or geopolitical developments.

Following this design, we will maintain a regular update cycle of every three weeks to keep the dataset aligned with real‑time safety risks.

---

Key Features

• 🌍 Real-world grounding: All instances are derived from publicly reported real-world events.
• ⏱ Temporal dynamics: The dataset is organized into time-indexed snapshots.
• ⚠️ Risk-triggered evaluation: Each instruction is explicitly linked to real-world risk triggers.
• 🔗 Compositional reasoning: Supports both single-event and multi-event risk scenarios.

---

Dataset Structure

WorldRiskBench is a time-evolving dataset composed of snapshots. Each file corresponds to a weekly snapshot constructed from a rolling window of recent real-world events.

---

Update Policy (Streaming Benchmark)

• The dataset is updated on a weekly basis
• Only the most recent two snapshots are retained in the public repository
• Older snapshots are removed to maintain a real-time evaluation setting

This design simulates realistic deployment conditions where models must reason over recently emerging risks, rather than static historical data.

---

Data Schema

Each record contains:

• risk_instruction (string): The user instruction to be evaluated
• risk_explanation (string): Explanation of why the scenario is risky
• risk_trigger (string): Real-world event that triggers the risk
• risk_type (string): Risk category
• url (list): source

---

Intended Use

WorldRiskBench is designed to evaluate world-dependent safety awareness, defined as:

The ability of models to recognize and reason about risks arising from interactions between user inputs and real-world events.

✅ Validated Use Cases

• Safety evaluation of LLMs and agent systems
• Benchmarking contextual risk reasoning under dynamic environments
• Evaluation of retrieval-augmented and tool-augmented systems

❌ Not Intended For

• Real-world decision-making systems
• Safety-critical deployment

---

Responsible AI Considerations

Data Limitations

WorldRiskBench is an evaluation-only benchmark designed to measure world-dependent safety risks and is not intended for model training, fine-tuning, or deployment. The dataset focuses exclusively on negative, safety-critical real-world events (e.g., disasters, recalls, incidents), limiting generalizability to general-purpose reasoning tasks. Data is primarily sourced from English-language news and structured databases, introducing geographic and linguistic coverage bias and underrepresenting low-resource regions. Although designed as a streaming benchmark, current releases consist of discrete temporal snapshots rather than continuous real-time updates. All evaluation instances are single-turn instructions and do not capture multi-turn or long-horizon agent behavior. The dataset includes LLM-generated instructions grounded in real events, which may introduce synthetic artifacts.

---

Data Biases

The dataset exhibits several forms of bias. First, selection bias arises from prioritizing high-impact, negative real-world events, leading to overrepresentation of crises and risk-heavy scenarios. Second, geographic and media bias is introduced through reliance on major English-language news sources and global reporting outlets, skewing coverage toward Western regions and widely reported events. Third, taxonomy-induced bias is present, as risk categories are artificially balanced during dataset construction and may not reflect real-world distributions. Fourth, generation bias is introduced by large language models used for instruction generation and annotation, which may reflect stylistic and cultural biases inherent in their training data. Fifth, compositional bias may arise in multi-event scenarios, where certain dominant risks are more salient than others, leading to uneven representation of subtle interactions. These biases may cause evaluated models to overfit to salient risk patterns and underperform on underrepresented regions or low-frequency event types.

---

Personal and Sensitive Information

The dataset does not intentionally include personally identifiable information (PII). However, it may contain indirect sensitive attributes derived from publicly available real-world events, including:

• Geographic locations (cities, regions, countries)
• Organizations (companies, institutions)
• Public figures (only when necessary for context)

All data is sourced from public reporting and processed to minimize privacy risks. However, due to the traceability of real-world events, residual sensitivity may remain.

---

Social Impact

WorldRiskBench enables systematic evaluation of AI safety under realistic, event-driven conditions, highlighting limitations in current models’ ability to handle dynamic and context-dependent risks.

However, potential risks include:

• Misuse for adversarial prompting or stress testing
• Over-reliance on benchmark scores as safety guarantees
• Bias toward negative world perceptions

To mitigate these risks, the dataset is:

• Evaluation-only
• Grounded in publicly verifiable events
• Free of intentionally harmful content

---

Important Notice

• ⚠️ This dataset contains references to real-world public events, which may include sensitive contexts.
• ⚠️ All data is publicly available and non-private.
• ⚠️ Some annotations are LLM-generated and may contain artifacts.

---