SentiClaw

SentiClaw 是一个 Python 安全中间件，在智能体收发消息的路径上依次做身份校验、注入清洗、出站敏感内容拦截、PII 脱敏、限流与回环检测、访问控制，并把事件写入 SQLite 审计库，可向指定告警通道推送威胁提醒。

supertechgod

热度 82.47

最近更新 2026-05-11 17:34:28

分类：安全依赖：代码执行 / 读写文件应用场景：工作流 / API 集成

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add senticlaw

Metadata

name: senticlaw
description: Runtime AI security for OpenClaw agents. Protects against prompt injection, identity spoofing, PII leakage, and runtime abuse. Drop-in 6-layer security middleware with SQLite audit logging, outbound content gating, and instant threat alerts. Use when: hardening your OpenClaw agent against external input attacks, protecting sensitive data in responses, or monitoring agent activity. NOT for: network/firewall security (this is AI-layer only).
metadata: {"openclaw":{"emoji":"🛡️","version":"1.0.0","author":"PHRAIMWORK LLC","url":"https://phraimwork.com","requires":{"python":">=3.10"}}}

SentiClaw — Runtime AI Security for OpenClaw

SentiClaw is a 6-layer security middleware that protects your OpenClaw agent from prompt injection, identity spoofing, data exfiltration, and runtime abuse.

Install

npx clawhub@latest install senticlaw
pip install ./skills/senticlaw

Quick Start

Add to your OpenClaw workspace (HEARTBEAT.md or any tool):

from senticlaw import SentiClaw

sc = SentiClaw(config={
    "owner_ids": {"discord": ["YOUR_DISCORD_USER_ID"]},
    "trusted_senders": {"discord": ["YOUR_DISCORD_USER_ID"]},
})

# Check inbound message
result = sc.check_inbound(text, sender_id=sender_id, channel="discord", session_id=session_id)
if not result.allowed:
    return result.block_message

# Run your agent logic here...
response = agent.respond(result.text)

# Check outbound response
safe = sc.check_outbound(response, session_id=session_id)
return safe.response

The 6 Layers

#	Layer	Protects Against
0	Identity	Spoofing, unauthorized access, name-claim attacks
1	Sanitizer	Prompt injection, jailbreaks, zero-width char attacks
2	Outbound Gate	API key leaks, internal IP exposure, system prompt leakage
3	Redactor	PII in responses (email, phone, SSN, credit cards)
4	Governance	Rate limiting, loop detection, spend caps
5	Access Control	Unsafe file paths, private URL access, tool abuse

Audit Log

All events are logged to SQLite (senticlaw_audit.db):

ALLOWED — clean message passed through
BLOCKED — message blocked by policy
INJECTION_ATTEMPT — prompt injection detected
SPOOFING_ATTEMPT — identity spoofing detected
OUTBOUND_BLOCKED — sensitive data in response blocked
RATE_LIMITED — sender exceeded volume limits
LOOP_DETECTED — repeated identical messages

Alert Integration

Wire up instant alerts to any channel OpenClaw supports:

sc = SentiClaw(config={
    "owner_ids": {"discord": ["YOUR_ID"]},
    "alert_channel":    "discord",    # discord | telegram | slack | whatsapp
    "alert_channel_id": "YOUR_CHANNEL_OR_CHAT_ID",
})

Any injection or spoofing attempt fires an immediate alert to your channel.

Configuration

from senticlaw import SentiClaw, SentiClawConfig

config = SentiClawConfig(
    owner_ids={"discord": ["YOUR_ID"]},
    trusted_senders={"discord": ["YOUR_ID"]},
    block_unknown_senders=False,
    redact_pii=True,
    redact_secrets=True,
    redaction_mode="mask",          # mask | remove | tokenize
    spend_cap_daily_usd=10.0,
    max_messages_per_hour=100,
    loop_threshold=3,
    outbound_block_api_keys=True,
    outbound_block_file_paths=True,
    alert_channel_id="",            # Discord channel ID for alerts
    audit_db_path="senticlaw_audit.db",
)

Running Tests

cd skills/senticlaw
python tests/run_tests.py

Built by PHRAIMWORK LLC · MIT License