← 返回 Skills 市场

Agent Safety Monitor

Name: Agent Safety Monitor
Author: evezart

作者 Evez666 · GitHub ↗ · v1.0.0 · MIT-0

cross-platform ⚠ suspicious

总下载

当前安装

版本数

在 OpenClaw 中安装

/install agent-safety-monitor

功能描述

Real-time AI agent safety monitoring, anomaly detection, and constraint enforcement. Use when building agent guardrails, detecting unsafe behaviors, enforcin...

使用说明 (SKILL.md)

Agent Safety Monitor

Monitor AI agent behavior in real-time, detect anomalies, and enforce safety constraints.

Quick Start

Run the monitor against an agent log stream:

python3 scripts/monitor.py --config safety_rules.yaml --input agent_logs/

Architecture

Agent Actions → Event Stream → Rule Engine → Alerts → Dashboard
                                   ↓
                              Anomaly Detector
                                   ↓
                              Auto-Pause/Kill

Safety Rule Types

Rate limits: Max N actions per minute/hour
Action allowlists: Only permit specific tool calls
Content filters: Block PII exfiltration, harmful outputs
Budget caps: Stop when cost exceeds threshold
Scope limits: Restrict file/API/network access
Behavioral anomalies: Flag unusual patterns (tool call frequency spikes, repetitive loops)

Rule Configuration

See references/rules-reference.md for the YAML schema.

Alert Channels

Console (default)
Webhook (Slack, Discord)
File log
Dashboard (HTTP server)

Dashboard

Launch the monitoring dashboard:

python3 scripts/dashboard.py --port 8080

安全使用建议

Review this before installing if you need real enforcement. The supplied code appears useful only for analyzing existing JSON logs and summarizing alerts; it should not be trusted to pause or kill agents, run a dashboard, or enforce scope limits without additional reviewed implementation. Do not provide wallets or sensitive credentials for this version.

能力标签

requires-walletrequires-sensitive-credentials

能力评估

⚠ Purpose & Capability

The stated purpose is real-time agent safety monitoring and constraint enforcement, but the supplied implementation appears to be an offline JSON log analyzer that records alerts and in-memory paused agent IDs rather than actually enforcing pause/kill or access controls.

ℹ Instruction Scope

The user-facing instructions are mostly scoped to running a local monitor script on user-provided logs, but they also advertise dashboard, webhook, and auto-enforcement features that are absent or unsupported in the provided artifacts.

ℹ Install Mechanism

There is no install spec, while the code imports external Python packages such as yaml and click and the SKILL references a missing dashboard script. This is mainly an operability/provenance gap, not evidence of malicious execution.

ℹ Credentials

Reading user-supplied agent logs and rule YAML is proportionate for a monitor, but registry capability signals mention wallet/sensitive credentials even though requirements and code do not justify them.

✓ Persistence & Privilege

The provided code shows no persistence, privilege escalation, credential use, network access, or background execution; event and alert state is in memory only.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install agent-safety-monitor
安装完成后，直接呼叫该 Skill 的名称或使用 /agent-safety-monitor 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v1.0.0

Initial release: real-time agent monitoring, anomaly detection, rate limits, budget caps, scope enforcement

元数据

Slug agent-safety-monitor

版本 1.0.0

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 1

常见问题