Automated Agent System Slashes Time to Build LLM Safety Benchmarks from Months to Days

Researchers have developed SafetyFlow, an automated system that dramatically accelerates the creation of safety benchmarks for large language models. The agent-flow platform can assemble a fully-formed safety evaluation dataset in just four days—eliminating the manual curation work that traditionally consumes months of researcher time and substantial resources.

The Problem with Current Approaches#

Testing whether large language models behave safely has become critical as their deployment expands. However, most existing safety benchmarks rely on painstaking human review to identify and label problematic model behaviors. This labor-intensive approach creates two significant bottlenecks: the sheer time investment required, and the tendency for manually curated datasets to contain overlapping questions and insufficient challenge variations. As a result, safety evaluations often miss edge cases and fail to probe the full range of model vulnerabilities.

How SafetyFlow Works#

Rather than human experts, SafetyFlow orchestrates a pipeline of seven specialized agents that collaborate to build safety benchmarks autonomously. Each agent handles a different phase of the benchmarking construction process. The system incorporates a suite of tools that let researchers control both the computational process and its costs while preserving human oversight at critical decision points. This hybrid approach ensures the final dataset reflects expert judgment without requiring manual annotation of every single query.

The SafetyFlowBench Dataset#

The system produced SafetyFlowBench, containing 23,446 distinct safety evaluation queries. Early analysis indicates the dataset exhibits low redundancy across queries and demonstrates strong ability to differentiate between models of varying safety levels—a critical feature for meaningful evaluation. The researchers validated their system by testing 49 advanced language models against SafetyFlowBench, confirming the benchmark's utility across a broad spectrum of state-of-the-art systems.

Broader Implications#

This work represents the first fully automated pipeline for constructing LLM safety benchmarks. By reducing the manual overhead that has historically bottlenecked safety research, SafetyFlow opens the door to more frequent, comprehensive, and rigorous evaluation of model safety as the field evolves.