HarmBench
HarmBench is an AI benchmark; this page records what it measures and traces its data, grader, and scores to primary sources.
Evaluates automated red-teaming attacks and model refusal defenses across standardized harmful behaviors.
- Address
- Online, Global
- Website
- harmbench.org/
- Tags
- Helpful, Viral, Science
- Category
- Safety
- Status
- Active
- Data licence
- MIT
- Data access
- open
- Grader
- mixed
- Organization
- Center for AI Safety
- Released
- 2024-02
- Leaderboard
- harmbench.org/
- Repository
- github.com/centerforaisafety/HarmBench
- Task count
- 510
- Top-score provenance
- unknown
- ABI family
- No
Questions
Where is HarmBench?
How do I contact HarmBench?
More benchmarks in Online
SafetyOnline
AdvBench
Measures whether adversarial suffixes induce aligned language models to produce targeted harmful behavior.
- Task count
- 520
AgentsOnline
AgentBench
Measures language-model agents across operating systems, databases, games, web shopping, and other interactive environments.
- Task count
- 8
SafetyOnline
AgentDojo
Evaluates agent task utility and resistance to prompt injection while agents use tools over untrusted data.
- Task count
- 97
ReasoningOnline
AGIEval
Tests general problem solving with questions from 20 public admission, qualification, and competition exams.
- Task count
- 20
AgentsOnline
AI2-THOR
Evaluates embodied agents in interactive simulated indoor scenes with object manipulation and navigation.
AgentsOnline
ALFWorld
Measures agents that learn high-level policies in text environments and execute the same household goals in embodied worlds.