首页 / 技能 / AI Safety Guard — 隐私外泄防护

📦 AI Safety Guard — 隐私外泄防护

v1.0.6

轻量级被动隐私守卫，专为 OpenClaw 设计，在 AI 发送或输出任何数据前智能拦截，防止用户敏感信息外泄，无需额外配置即可静默守护。

5· 298·0 当前·0 累计
by @andreqingyuwu (Andre Wu)
安全AI模型访问API工具
下载技能包
最后更新
2026/4/22
0
安全扫描
VirusTotal
无害
查看报告
OpenClaw
可疑
medium confidence
The skill's stated purpose (preventing data leaks) matches being an instruction-only privacy guard, but the SKILL.md contains contradictory rules about sending credentials and confirmations that make its runtime behaviour ambiguous and potentially risky.
评估建议
This skill is a rulebook the agent would follow to decide whether to send data externally. That design is reasonable, but the SKILL.md contains contradictory guidance in safety-critical places — notably around credentials and phishing handling. Before installing or enabling it: - Ask the author to resolve contradictions (explicit precedence): clarify whether 'never send raw passwords/API keys' ever admits an exception for 'user-requested logins', and whether 'silently execute' means the user i...
详细分析 ▾
✓ 用途与能力
The name/description (a passive privacy guard) align with an instruction-only skill that tells the agent how to decide about external transmissions. No unrelated environment variables, binaries, or install steps are requested, which is proportionate to the claimed purpose.
⚠ 指令范围
The SKILL.md defines high-level decision logic for blocking, anonymizing, or allowing transmissions, but contains contradictory and ambiguous directives: e.g. it lists 'passwords, API keys, bearer tokens' as 'No useful anonymized form (never send raw)', yet Scenario 1 instructs: 'Log into Gmail... → SILENTLY EXECUTE → Execute the login, send credential to Gmail.' There are also conflicts about user interaction: 'Do not offer options, do not ask for confirmation' vs. 'Wait for the user to either correct the destination or explicitly confirm.' Terms like 'silently execute' vs. 'notify' are used inconsistently. These contradictions make it unclear what the agent should actually do in key cases (credentials, suspected phishing, background transmissions). The skill also instructs the agent to look at recent user messages and whether data was found 'in a file' but does not bound what files or contexts to inspect; that grants broad discretion to the agent in the absence of stricter rules.
✓ 安装机制
No install spec and no code files (instruction-only). This is the lowest-risk distribution model and consistent with a policy-style skill that provides agent guidance rather than executable artifacts.
✓ 凭证需求
The skill requires no environment variables, credentials, or config paths, which is proportionate. The instructions reference domains and destinations but do not request external credentials or keys from the host.
✓ 持久化与权限
The skill does not request always:true, does not declare install-time writes, and is user-invocable only. It does not request persistent presence or modification of other skills' configs.
安全有层次，运行前请审查代码。

运行时依赖

无特殊依赖

版本

latestv1.0.62026/3/15

Summary: This update changes the AI privacy guard from a silent/passive protector to an informative one, providing user notifications for all security actions without interrupting workflow. - The guard now briefly notifies users of any data protection actions taken (execution, anonymization, cancellation, or block). - All decisions (execution, anonymization, cancellation, phishing block) are now explicitly communicated to the user in a non-intrusive way. - Phishing attempts are blocked and the user receives a one-time warning. - The principle and scenario descriptions have been updated to reflect the new informative notification approach. - Workflow remains non-disruptive—no confirmations or prompts are required from users.

● 无害

安装命令

点击复制
官方npx clawhub@latest install ai-safety-guard
镜像加速npx clawhub@latest install ai-safety-guard --registry https://cn.longxiaskill.com
数据来源：ClawHub ↗ · 中文优化：龙虾技能库