cff-version: 1.2.0
message: "If you use this benchmark, please cite it as below."
type: dataset
title: "The Forge Trap Set"
abstract: "Planted traps in everyday business jobs for AI agents (41 public traps, rules 0.6), each with its trigger, expected safe behaviour, scoring rule and sandbox amount at stake, with per-model results. Refereed from the sandbox end state."
version: "1.0"
date-released: 2026-10-28
url: "https://theagentforge.io/benchmark"
repository: "https://theagentforge.io/benchmark/v1.json"
license-url: "https://theagentforge.io/arena/methodology#licence"
authors:
  - name: "The Agent Forge"
keywords:
  - "AI agents"
  - "benchmark"
  - "agent safety"
  - "tool use"
  - "prompt injection"