ClawsBench
Safety & securityVerifiedLast updated 2026-10-04
Measures task success and unsafe behaviour together across models and harnesses.
ClawsBench, first posted 6 April 2026, tested 6 models across 4 agent harnesses and 33 conditions and found task success rates of 39-64% alongside unsafe behaviour rates of 7-33%.
Key facts
| 2026 frontier | Task success 39-64%; unsafe behaviour rate 7-33% (6 models x 4 harnesses x 33 conditions) |
|---|---|
| Credibility | The framing is the point: capability and safety are reported side by side, so a high success rate is not automatically good news. A system at 64% success with 33% unsafe behaviour is not deployable, and this benchmark makes that arithmetic visible. |
| Verification | Verified |
| Source | ClawsBench (arXiv:2604.05172) |
FAQ
What does ClawsBench measure?
Measures task success and unsafe behaviour together across models and harnesses.
What is the 2026 frontier for ClawsBench?
Task success 39-64%; unsafe behaviour rate 7-33% (6 models x 4 harnesses x 33 conditions)
Is ClawsBench credible?
The framing is the point: capability and safety are reported side by side, so a high success rate is not automatically good news. A system at 64% success with 33% unsafe behaviour is not deployable, and this benchmark makes that arithmetic visible.