/

← All benchmarks

AgentHarm

safety · agentic-misuse↗ Quelle

53Evidence

not yet measured

What it measures

Ob ein Agent schaedliche Auftraege ausfuehrt, wenn Werkzeuge bereitstehen.

What it does not measure

Nuetzlichkeit, Ueberblockierung.

How the score is composed

Currency50

Discrimination50

Headroom50

Contamination safety33

Gaming resistance100

Methodological rigor80

Reproducibility20

Details

Benchmaxxing risk: low
Difficulty: medium
Public test set: yes
Human-verified: yes
LLM-judged: no
Skala: 0–100 % refusal (höher besser)

Hand-maintained — no automatic adapter yet.

Best measured models

No results ingested yet.