Skip to content

Dataset · harness-hub

Agent tool call safety benchmark

harness-hub/safeagents-judge-bench

Real English AI agent trajectories for evaluating the safety of their actions.

What's inside

A small English dataset (fewer than 1,000 entries) with real tool call logs recorded during coding agent operations. Examples include agent attempts to access operator accounts and storage, with real access keys hidden.

What you can do with it

  • Testing judge models and safety filters for AI agents
  • Evaluating coding agent vulnerabilities to unauthorized data access
Tasks
Text classification
Languages
English
Size
up to 1K rows
Downloads this month
2.4K
Likes
30

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("harness-hub/safeagents-judge-bench")