Skip to content

Dataset · ai-safety-institute

AgentHarm benchmark for LLM agent harmfulness

ai-safety-institute/AgentHarm

Small text benchmark for measuring the harmfulness of LLM agent actions.

What's inside

AgentHarm is a benchmark for measuring LLM agent harmfulness; the card lists authors from Gray Swan AI, the UK AI Safety Institute, and universities. The dataset is text-based and very small—under 1K examples. The language is not specified in the metadata.

What you can do with it

  • Evaluating LLM agent safety
  • Red-teaming agentic systems
  • Checking guardrails and filters
Size
up to 1K rows
Downloads this month
16K
Likes
68

How to load it

from datasets import load_dataset

ds = load_dataset("ai-safety-institute/AgentHarm")