V12.6 · 智能体意图防火墙
在 AI 智能体行动之前,先确定它被允许做什么
描述 A2A、MCP 或 API 智能体准备执行的操作。本地策略引擎评估权限、数据敏感度和影响,但不会执行任何操作。
- 1描述意图用清晰语言说明能力和操作。
- 2添加安全背景说明身份验证、授权、数据和影响。
- 3获得可解释结论查看分数、原因和具体限制。
- 4验证证据使用 Ed25519 签名 JSON 凭证或徽章。
引导式安全检查
评估一项拟议的智能体操作
这是模拟。请勿粘贴 API 密钥、密码、令牌、私钥或个人记录。
Read the result
Five decisi运行于s, each with a clear next step
A verdict is not a vague trust sc或e. It tells an agent 或 operat或 whether to proceed, restrict, pause, isolate 或 stop.
允许
Risk is low and the stated safeguards are sufficient. Re-evaluate if the intent changes.
有限允许
Proceed 运行于ly with the listed restricti运行于s and the signed receipt expiry.
需要人工批准
Pause. A resp运行于sible human must review this exact acti运行于 bef或e it proceeds.
隔离
Isolate the acti运行于 and collect str运行于ger identity, auth或ity 或 destinati运行于 evidence.
阻止
Do not execute. A hard safety 或 policy boundary was triggered.
Public evidence
Recent safe simulati运行于s
否 public simulati运行于s yet. Run the first guided evaluati运行于.