What this risk is
Tool-using agents can call the right tool for the wrong file, state, identity, or permission boundary.
EvalDuel growth page
Tool-Use Safety Eval
Agent 红队试炼场
Your AI agent looks smart. EvalDuel finds how it breaks.
Why normal evals miss it
Static safety prompts do not prove the executed tool boundary is protected under pressure.
How EvalDuel tests it
EvalDuel makes agents defend virtual workspaces, files, and tool policies while hidden validators score both safety and utility.
回放 case
让 EvalDuel 攻击你的 Agent
把这个失败模式变成针对你自己自主 Agent 的可复盘 pilot。