← 返回 Skills 市场
Prompt Guard
作者
seojoonkim
· GitHub ↗
· v3.6.2
13386
总下载
57
收藏
0
当前安装
17
版本数
在 OpenClaw 中安装
/install prompt-guard
功能描述
650+ pattern AI agent security defense covering prompt injection, supply chain injection, memory poisoning, action gate bypass, unicode steganography, cascad...
安全使用建议
Install only if you are comfortable with default outbound calls for remote pattern updates and HiveFence threat reporting. For sensitive or regulated environments, disable API and HiveFence/network reporting before first use, review logging settings so prompt previews are not stored unexpectedly, and avoid running the Clawdbot audit --fix path unless you intend local permission changes.
功能分析
Type: OpenClaw Skill
Name: prompt-guard
Version: 3.6.2
The prompt-guard skill bundle is a comprehensive security library for AI agents, providing over 650 detection patterns for prompt injection, data exfiltration, and tool abuse. It features a multi-layered defense-in-depth architecture including text normalization, multi-encoding decoders (decoder.py), and an enterprise-grade DLP system (output.py) for redacting sensitive credentials. The bundle includes an optional API client (api_client.py) and a distributed threat intelligence client (hivefence.py) that report anonymized threat metadata (hashes) to external endpoints (pg-secure-api.vercel.app and hivefence-api.seojoon-kim.workers.dev). A system audit utility (audit.py) is also provided to check for common security misconfigurations. The code is well-structured, includes extensive regression tests, and demonstrates clear intent to protect agents rather than attack them.
能力评估
Purpose & Capability
The main prompt-injection, DLP, normalization, decoding, logging, and pattern-scanning capabilities fit a security skill, but runtime code also defaults to API pattern fetching and HiveFence HIGH+ detection reporting, and ships Clawdbot host-audit code that is broader than the core prompt-guard purpose.
Instruction Scope
Documentation is inconsistent: it advertises offline or optional network operation in some places while SKILL.md, README.md, architecture docs, config, and engine defaults enable API access by default; HiveFence auto-reporting is less clearly disclosed than the API pattern fetch path.
Install Mechanism
The Python package metadata and dependencies are ordinary and there are no install-time hooks or obfuscated installers observed, but the package includes scripts and audit utilities beyond the primary library surface.
Credentials
Default network egress to the pattern API and HiveFence expands the trust boundary for a tool that may inspect sensitive prompts and security events; local logging to memory/security-log files is purpose-aligned but should be made explicit to users.
Persistence & Privilege
No background persistence was found, but the skill writes security logs by default, the HiveFence client can cache data under ~/.clawdbot, and the audit utility can chmod local Clawdbot paths when run with --fix; these are not automatically destructive but require user awareness.
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 prompt-guard - 安装完成后，直接呼叫该 Skill 的名称或使用
/prompt-guard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
v3.6.2
No code or documentation changes detected in this release.
- Version number updated from 3.6.0 to 3.6.2.
- No functional or documentation changes present.
v3.6.1
No changes in this release.
- Version number updated only.
- No file changes detected.
- No new features, fixes, or updates introduced.
v3.6.0
**v3.6.0 expands coverage to 650+ patterns with new ClawSecurity-aligned detections and attack categories.**
- Added 50+ new detection patterns, including ClawHavoc supply chain signatures, cloud credentials exfiltration, and code exfiltration defense.
- Introduced detection for multi-turn manipulation, authority escalation (e.g., emergency override, sudo grant), and PII output such as SSN and credit cards.
- Enhanced protection against config drift, large data dumps, SQL injection via tool parameters, and path traversal attacks.
- High and medium tiers expanded with new checks for financial data, cross-session attacks, and tool/agent parameter abuse.
- Pattern set updated: now covers prompt injection, supply chain threats, memory poisoning, unicode steganography, cascade amplification, and more across 12 security categories and 10 languages.
v3.5.0
v3.4.0/3.5.0: Typo-based evasion detection + TieredPatternLoader fix (PR #10 by @matthew-a-gordon). 14 new regression tests. Drop-in LLM prompt injection defense.
v3.4.0
v3.4.0: AI Recommendation Poisoning, Calendar Injection, PAP Social Engineering
v3.3.0
**v3.3.0 adds optional API support with early-access and premium pattern tiers.**
- Introduced API client for early-access and premium pattern updates (optional; uses built-in beta key, can be disabled).
- Bundled pattern set expanded to 577+; now includes advanced "skill weaponization" patterns for deeper threat coverage.
- Switchable between fully offline (no API requests) and API-enhanced detection via config or environment variable.
- Improved documentation: clarified API usage, pattern tiers, config, and CLI; security categories and feature set updated.
- New and updated tests for typo evasion and API behavior.
- Internal code updates in engine, scanner, and patterns to support API integration and expanded tier logic.
v3.1.0
Token Optimization: 70% reduction via tiered loading, 90% cache savings, SKILL.md 65% smaller
v2.6.1
# prompt-guard v2.6.1 Changelog
- Updated changelog and documentation.
- Minor adjustments in scripts/detect.py (details not specified in input).
v2.5.3
**prompt-guard v2.6.0 – Major update: HiveFence distributed threat intelligence and new real-world defenses**
- Integrated with HiveFence: agents share and receive new attack patterns via collective defense network.
- New CLI tools for reporting, voting, and syncing threat patterns with HiveFence.
- Additional defenses against social engineering, including single-approval expansion, credential path harvest, and security bypass coaching.
- Owner-only restrictions now enforced for sensitive commands in DMs as well as group chats.
- ARCHITECTURE.md added; major documentation updates.
v2.5.2
Moltbook attack collection: BRC-20 JSON injection, guardrail bypass, agent sovereignty manipulation, CALL TO ACTION detection
v2.5.1
prompt-guard v2.5.1
- Added critical detection for LLM system prompt mimicry (e.g. fake Claude/Anthropic/GPT/LLM tokens, tags, and famous jailbreak markers).
- Blocks attacks attempting to poison session context via `<claude_*>`, `<|im_start|>`, `[INST]`, `GODMODE`, `DAN`, `JAILBREAK`, leetspeak variants, and similar prompts.
- Expanded detection coverage for real-world prompt injection and context poisoning exploits.
- New documentation: Added SECURITY.md and blog post explaining new defenses.
- Updated and reorganized SKILL.md for improved clarity.
v2.3.0
Fix: clarify loopback vs webhook mode
v2.2.1
v2.2.1: Enhanced README with threat scenarios, changelog, version badges
v2.2.0
v2.2: Secret protection (blocks token/config requests in EN/KO/JA/ZH), security audit script, infrastructure hardening guide, SSH/gateway/browser security checks
v2.1.0
v2.1: Full English documentation, improved config examples, comprehensive testing guide
v2.0.0
v2.0: Multi-language support (KO/JA/ZH), severity scoring, homoglyph detection, rate limiting, security log analyzer, configurable sensitivity
v1.0.0
Initial release: prompt injection defense for group chats
元数据
常见问题
Prompt Guard 是什么？
650+ pattern AI agent security defense covering prompt injection, supply chain injection, memory poisoning, action gate bypass, unicode steganography, cascad... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 13386 次。
如何安装 Prompt Guard？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install prompt-guard」即可一键安装，无需额外配置。
Prompt Guard 是免费的吗？
是的，Prompt Guard 完全免费（开源免费），可自由下载、安装和使用。
Prompt Guard 支持哪些平台？
Prompt Guard 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Prompt Guard？
由 seojoonkim（@seojoonkim）开发并维护，当前版本 v3.6.2。
推荐 Skills