← 返回 Skills 市场
403
总下载
1
收藏
1
当前安装
12
版本数
在 OpenClaw 中安装
/install openclaw360
功能描述
Runtime security skill for AI agents — prompt injection detection, tool call authorization, sensitive data leak prevention, skill security scanning, and one-...
安全使用建议
Things to consider before installing:
- The skill bundle contains only instructions; installing requires running pip to fetch code from GitHub. Review the referenced repo and the exact pinned commit (5fd69db) before installing to ensure the code matches expectations.
- The tool will create ~/.openclaw360/, generate an Ed25519 identity key (0600), keep audit logs and backups there — confirm you are comfortable storing these artifacts locally and check permissions.
- The documented behavior allows a degraded fallback (continue operation) if checks fail; if you need hard enforcement, confirm or modify that policy in the tool's config.
- The agent will not autonomously invoke this skill (disable-model-invocation: true), but when you enable the proactive protection mode the agent will call the local CLI before inputs/tools/outputs — be cautious when asking it to scan arbitrary directories (it reads files you point it at).
- If you cannot or will not review the upstream source, treat the pip-install-from-GitHub step as a moderate trust decision and consider installing into an isolated virtualenv before granting broader access.
Confidence note: assessment is 'benign' but confidence is medium because the skill bundle omits code — verifying network/no-network runtime behavior and exact implementation requires inspecting the upstream repository at the pinned commit.
功能分析
Type: OpenClaw Skill
Name: openclaw360
Version: 0.1.11
The openclaw360 skill is a comprehensive security utility designed to provide runtime protection for AI agents, including prompt injection detection, tool call authorization, and sensitive data masking. It operates locally, stores data in a dedicated directory (~/.openclaw360/), and uses Ed25519 signatures for log and backup integrity. The instructions guide the agent to implement a 'security middleware' pattern by checking inputs and tool calls via a CLI tool installed from a pinned GitHub commit (https://github.com/milu-ai/openclaw360), which is consistent with its stated purpose and lacks evidence of malicious intent.
能力评估
Purpose & Capability
Name/description (runtime security, prompt-injection/tool/output checks, skill scanning, backup/restore) align with the declared requirements: only python3 and a dedicated ~/.openclaw360 config tree. The metadata.install commands (pip from the repo) are logically required to obtain the tool.
Instruction Scope
SKILL.md instructs the agent to invoke a local CLI (openclaw360) for prompt/tool/output checks and to scan user-specified Skill directories. Commands read user-supplied text, SKILL.md files in paths the user provides, and audit logs under ~/.openclaw360/. This stays within the stated security purpose, but the instructions also explicitly say the agent will continue execution if checks fail (degraded mode), which is a design choice users should be aware of.
Install Mechanism
Installation is via pip from a pinned GitHub commit (git+https://github.com/milu-ai/openclaw360.git@5fd69db) or a venv variant. Using a pinned commit is better than an unpinned install, but fetching and executing code from a remote repo is moderate risk. The skill bundle itself contains no code to audit, so verification requires the external repo review.
Credentials
No environment variables or unrelated credentials are requested. The skill only needs and will create/read files under ~/.openclaw360/ (config, identity keys, audit, backups). Those local config paths are proportionate to the described functionality.
Persistence & Privilege
disable-model-invocation: true and always: false are set (agent won't be auto-triggered). The skill only requests to create and manage its own ~/.openclaw360/ tree (including an Ed25519 identity key with 0600), which is consistent with local audit/backup functionality and not an elevated or system-wide privilege.
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 openclaw360 - 安装完成后，直接呼叫该 Skill 的名称或使用
/openclaw360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
v0.1.11
openclaw360 v0.1.11
- Added support for new backup-list trigger type: pre_restore.
- Enhanced backup-clean and backup-clean [--dry-run] CLI options for safer backup management.
- Updated documentation to reflect these options and usage.
- No changes to logic or behaviors outside of improved backup operations and command documentation.
v0.1.10
**Clarifies active protection logic and invocation policy.**
- Updated documentation to explicitly clarify the relationship between `disable-model-invocation: true`, user-initiated protection mode, and when CLI-based security checks are allowed.
- Added explanation that active protection and CLI checks require explicit user request, and this does not conflict with the disable-model-invocation setting.
- Enhanced downgrade behavior description: clarified that failed security checks notify the user but do not block agent operation, and explained the security design trade-off.
- Provided clarifications regarding agent/skilled behavior when active protection is enabled at user request.
- No changes to code or features; documentation and instruction clarifications only.
v0.1.9
**openclaw360 v0.1.9 changelog**
- Added section detailing "主动防护模式" (proactive runtime protection), clarifying when and how to invoke security checks at input, tool call, and output stages.
- Describes new CLI invocation patterns for check-prompt, check-tool, and check-output, along with decision handling (allow/block/confirm) and error fallback behavior.
- Emphasizes agent-level integration for runtime checks without relying on platform hooks.
- Clarifies downgrade/fallback mechanisms if security CLI commands fail or are unavailable.
- No changes to code execution, permissions, or installation steps; documentation update only.
v0.1.8
- Updated skill version from 0.1.6 to 0.1.8.
- Promoted the "install" field to the top-level frontmatter, providing clear installation commands.
- Changed scan-skills execution and report rules from strict ("必须严格遵守"/"强制") to recommendations ("建议遵循"/"推荐"), making reporting templates and rules more flexible.
- Minor edits to rules language regarding scan and reporting steps to soften mandatory instructions and allow more leeway in result formatting and response phrasing.
- No changes to core functionality or security logic.
v0.1.7
**Added one-click backup and restore feature for security configuration and audit data.**
- Introduced new `backup`, `restore`, `backup-list`, `backup-verify`, and `backup-clean` commands for managing secure backups.
- Updated permissions and data handling to include `~/.openclaw360/backups/` directory and signed backup manifests.
- Instructions and rules now specify atomic backup behavior and smart clean-up logic.
- Security scanning/reporting rules clarified and strengthened, with stricter enforcement of language, output templates, and reporting requirements.
- Documentation expanded to reflect enhanced data protection in backup and restore workflows.
v0.1.6
- Clarified that all post-install commands run locally with no network requests, and that network behavior can be reviewed via the open-source repo.
- Updated scan-skills instructions: clarified default/typical scan directories and explicitly stated sensitive directories are not accessed.
- Enhanced documentation around Skill scan paths to distinguish between openclaw360 and OpenClaw platform directories.
- No functional or code changes; documentation only.
v0.1.5
openclaw360 0.1.5
- Updated SKILL.md to set `disable-model-invocation: true` as a top-level frontmatter field (previously under metadata).
- Updated pinned commit hash in install instructions for both pip and venv methods.
- Minor documentation clarifications; no code or functionality changes.
v0.1.4
- Updated pinned commit hash in pip install command for improved code consistency and integrity.
- Clarified permission, reading, and writing behaviors in the SKILL.md for better transparency.
- Refined data handling and network access descriptions; now explicitly lists which files and directories are accessed or excluded.
- Simplified instructions and feature descriptions, making the documentation more concise and user-friendly.
- Removed redundant details from usage and scan-report rules, while retaining core security principles.
v0.1.3
**Summary:** Improves security transparency and operational safety, clarifies "read-only" design, and adds install instructions to metadata.
- Declares all functionality as pure read-only analysis (no shell commands, file modification, or outbound network requests).
- Expands and clarifies Permissions, Data Handling, and Network Access sections; affirms no telemetry or background activity.
- Introduces install instructions directly within metadata.install for user reference and reproducibility.
- Enhances security guarantees: MIT license, source-verifiable, no code execution or subprocess calls, and no external data flows.
- Clarifies doc- and script-based sensitive data distinction in audit/scanning reports.
- Updates example commands, scan report display rules, and ALL operational instructions for consistency with safer usage.
v0.1.2
openclaw360 v0.1.2 changelog:
- Expanded and clarified skill security scanning rules, including strict reply and language requirements.
- Detailed requirements and templates for scan report structure, with a two-stage (summary/detail) presentation process.
- Added explicit instructions to run scan-skills only once per scan and to wait for full output before replying.
- Further specified prompt injection detection methods, risk formulas, and tool-call authorization process.
- Improved documentation for DLP coverage, audit logging, and user communication practices.
v0.1.1
- Added explicit requirement for user confirmation before installation, detailing created files and directories.
- Enhanced security: now requires all installs to use a pinned commit hash (not just a tag) for source code integrity.
- Clarified that all data remains local; Ed25519 identity keys are only used for audit log signing.
- Updated instructions: network access is only allowed for install (once) and user-triggered rules update—no background or automatic network requests.
- Improved documentation of new command-line options (`check-prompt`, `check-tool`, `check-output`) and mandatory local directories.
- Restricted model invocation by adding `disable-model-invocation: true` in metadata.
v0.1.0
Initial release of openclaw360
- Provides runtime security protection for AI agents with prompt injection detection, tool call authorization, sensitive data leak prevention, and skill security scanning.
- All data is processed locally; audit logs store only SHA-256 hashes; no data is sent to external servers.
- Installation workflow supports both direct pip install and venv fallback for compatibility.
- Skill security scanning supports static analysis, JSON output, and minimum score filters.
- Does not require sudo or access to sensitive system directories; needs Python 3.10+.
元数据
常见问题
Openclaw360 是什么？
Runtime security skill for AI agents — prompt injection detection, tool call authorization, sensitive data leak prevention, skill security scanning, and one-... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 403 次。
如何安装 Openclaw360？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install openclaw360」即可一键安装，无需额外配置。
Openclaw360 是免费的吗？
是的，Openclaw360 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Openclaw360 支持哪些平台？
Openclaw360 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 Openclaw360？
由 WalkerX（@326668808）开发并维护，当前版本 v0.1.11。
推荐 Skills